On this page

Summary
Why this question
Scope and methods
The landscape
Theme 1 — The concept and its theoretical foundations
Theme 2 — Self-training: bootstrapping from one’s own outputs
Theme 3 — Self-correction and self-refinement
Theme 4 — Self-rewarding and alignment without human labels
Theme 5 — Self-play
Theme 6 — Synthetic-data loops and model collapse
Theme 7 — Weak-to-strong generalization
Theme 8 — Agentic self-evolution and open-ended improvement
Theme 9 — Safety, stability, and benchmarks of recursive improvement
Where the evidence disagrees
Gaps and open questions
Confidence and limitations
Evidence table
References

Recursive self-improvement in large language models

How do large language models improve themselves recursively, through which mechanisms, with what empirical results, and under what limits?

Updated: 10 Aug 2026
Sources: 156
Years: 2006–2026
Confidence

https://reviews.lewiswon.me/reviews/recursive-self-improvement-llms/ · Updated 10 Aug 2026

Download Markdown

recursive self-improvement self-training self-correction self-play RLAIF model collapse weak-to-strong generalization self-evolving agents intelligence explosion

How this review was made

Databases: OpenAlex, arXiv, Semantic Scholar, Crossref, DBLP
Queries (literal): recursive self-improvement; recursive self improvement; recursive self-improvement language model; self-improvement large language model; self-improving language models; self-evolution large language models; self-evolving agents; self-training large language models; self-play large language models; self-rewarding language models; self-rewarding; self-refine language models; self-correction large language models; self-correct reasoning language models; self-alignment language models; RLAIF reinforcement learning from AI feedback; constitutional artificial intelligence; weak-to-strong generalization; model collapse generative; self-consuming generative models; curse of recursion synthetic data; synthetic data large language models training; expert iteration; iterated distillation and amplification; Godel machine self-improvement; intelligence explosion; self-taught reasoner; self-taught optimizer; self-verification language models; self-play preference optimization; self-improving AI; iterative self-training; self-distillation; machine self-improvement
Search last run: 2026-08-10
Screening: 156 sources used · 2006–2026 · deep review

Summary

The short version

This review asks how large language models improve themselves recursively: through which mechanisms, with what empirical results, and under what limits. Across 156 sources spanning 2006–2026, the evidence shows that single-loop self-improvement is real, repeated, and often substantial — models trained, refined, or rewarded on their own outputs routinely gain 10–30 percentage points on reasoning and instruction-following benchmarks. The evidence is equally clear that the recursion is fragile: every family of methods has a documented failure mode (reward hacking, self-correction artifacts, model collapse in synthetic-data loops, capability forgetting), and a 2025–26 wave of papers is converging on the same conclusion — that improvement from one’s own signal is bounded by the quality and diversity of that signal unless an external anchor (verifiable feedback, fresh data, human judgment) is injected. Confidence is moderate: the field is young, dominated by preprint, author-run benchmarks, and a small set of labs, and the headline numbers have rarely been independently replicated. The single most important caveat is that most reported gains are measured on the authors’ own evaluation setups, and several landmark results (self-correction, self-rewarding, model-collapse avoidance) are actively contested.

Why this question

“Recursive self-improvement” (RSI) is the oldest and most consequential hypothesis in AI: if an AI system can improve its own capability, the improvement can compound, and the result is an intelligence explosion — a transition from human-level to vastly superhuman AI within a short window, with all the risk that entails. For most of the field’s history the idea lived in philosophy and safety theory, because no system could actually do it. That changed in 2022–2024: language models turned out to be trainable on their own outputs (self-training, self-play, self-rewarding), and a flood of papers now claim exactly the kind of capability loop the theorists described. This is the first literature in which RSI is an empirical research program rather than a thought experiment.

What turns on the answer is unusually large. If recursive improvement robustly compounds, frontier-lab economics and AI-risk assessments change (the “intelligence flywheel” arguments); if it reliably stalls or collapses without external anchors, then the intelligence-explosion scenario needs a different mechanism than anything demonstrated, and safety work can focus on more concrete problems. The empirical literature now speaks to both possibilities, and this review maps what it actually shows.

Scope and methods

Question. How do large language models improve themselves recursively — through what mechanisms, with what empirical results, and under what limits?

Inclusion criteria. Works about self-improvement of neural language models or their direct precursors: self-training and bootstrapped reasoning (STaR, expert iteration), self-correction and self-refinement, self-rewarding and alignment from AI feedback (RLAIF), self-play, synthetic-data training loops and model collapse, weak-to-strong generalization, agentic self-evolution, and the theoretical literature on recursive self-improvement (Gödel machines, seed AI, intelligence explosion). English-language, published or arXiv, 2006–2026. Both empirical and theoretical work included.

Exclusion criteria. Domain applications of self-training outside LLMs (medical imaging, speech recognition, gene prediction) unless they are the canonical precursor (Noisy Student); self-distillation for model compression unless it addresses capability dynamics; prompt-engineering work without a self-improvement loop; AI-safety policy essays without technical content; pure philosophy of mind. Two sources were dropped at screening because no open copy could be retrieved in-session (EURISKO, Lenat 1983 — paywalled, no open version; and an SSRN-only survey), and two theory preprints hosted only on bot-walled platforms (Zenodo/TechRxiv-adjacent).

Search and screening. 34 concept queries across OpenAlex, arXiv, and Semantic Scholar (Semantic Scholar was rate-limited and contributed little); plus exact-title resolution ladders (Crossref, arXiv API, abs-page verification) for landmark papers the keyword search missed, and snowballing from the citation graph of key anchors. Retrieved 1,395 unique records; 495 passed a title-level relevance gate; 156 were selected after full-record curation. Every arXiv ID was verified against its abstract page (title similarity ≥ 0.85); every DOI was verified via Crossref; three fuzzy-matched DOIs were caught and corrected. 135 sources were read in full text; 21 were abstract-only. The literal query strings are recorded in the frontmatter.

Coverage notes. The review deliberately covers theory and history lightly — the pre-LLM RSI literature is included for context, not exhaustively. The 2026 literature is captured through early August 2026. Semantic Scholar’s outage meant its (largely redundant) coverage was thinner than planned.

The landscape

The shape of this literature is unusual: a long, slow theoretical tail (2006–2021), a sudden empirical burst (2022–2024) built on a small set of landmark papers, and a 2025–26 explosion of variants, benchmarks, and safety critiques — roughly half of the 156 sources here are from 2025–26, and the RSI-specific vocabulary (“self-evolution”, “self-rewarding”, “recursive self-improvement”) only crystallized in that window. The field is young even by ML standards: most results are arXiv preprints, most benchmarks are author-run, and a few labs dominate the landmark papers (Stanford/FAIR for STaR and its descendants, Meta for self-rewarding and synthetic-data collapse, DeepMind for self-play and open-endedness). One benchmark family dominates measurement: GSM8K, MATH, and HumanEval appear in most empirical papers, which makes cross-paper comparison possible but also means the evidence is concentrated on math and code, where verification is cheap — a selection effect that runs through the whole review. The theoretical literature, by contrast, is a genuine multidisciplinary conversation: computer science (Gödel machines, bounded RSI), philosophy (intelligence explosion), economics (compute bottlenecks), and empirical ML meet only rarely, and one of the striking findings is how little the empirical and theoretical strands cite each other.

Theme 1 — The concept and its theoretical foundations

The idea of a system that improves itself has a formal core that predates LLMs by decades. Schmidhuber’s Gödel machines are the canonical formalization: a self-referential problem solver that can rewrite its own code, with a proof searcher that only allows rewrites it can prove useful — the Global Optimality Theorem shows any such provably useful rewrite is globally optimal, with no local maxima ¹⁵⁶. The framework’s successors have kept the aspiration while relaxing the proof burden: Nivel et al.’s bounded recursive self-improvement blueprint (autocatalysis, endogeny, reflectivity) trades provable optimality for an architectural account of how a system could bootstrap its own learning algorithm ¹, and Steunebrink et al. argue that most treatments consider only pairs of (agent, modification candidate) and miss the broader design space ⁵. Yampolskiy provided the field’s working definitions — weak vs strong RSI — and a survey of self-improving software types, while arguing that computational limits constrain the recursion ²³. Wang et al. give a formal model in which an RSI procedure reaches a global optimum in logarithmic expected steps relative to search-space size, in simulation ⁴. A 2026 theoretical note connects sustainable RSI to an “introspection threshold” analogous to von Neumann’s complexity threshold, arguing that introspective programs provably exist (via Kleene’s recursion theorem) but current LLMs sit below the threshold ¹²⁵.

The philosophical lineage runs through the intelligence explosion. The canonical arguments — that a human-level AI could bootstrap itself to vastly superhuman capability — are presented and assessed by Muehlhauser and Salamon (who find the case substantial but the evidence base thin) ⁷, Chalmers ⁸, and Jebari, who argues that because human “techne” (problem-solving via artifacts) varies more across individuals than raw intelligence, a machine techne surpassing human techne would likely produce a prolonged transition rather than a sudden explosion ⁹. The deepest formal limit argument in this lineage is Gödel’s dichotomy, analyzed by Feferman: either the human mind infinitely surpasses any finite machine, or there exist absolutely unsolvable problems no formal system can decide — either horn constrains what a recursively self-improving formal system can provably achieve ⁶. The modern descendants of this debate are empirical-adjacent: MacAskill’s preparedness analysis of an AI-accelerated research explosion ¹¹; Whitfill’s compute-bottleneck model, which finds the answer depends on whether compute and labor are substitutes or complements — his own two specifications give wildly different explosion probabilities (σ = 2.58 vs −0.10) ¹⁰; and a survey of 25 AI researchers in which 20 of 25 called automating AI research one of the most severe and urgent risks, while 17 of 25 still expected the transition to be gradual or survivable ¹³. Evans et al. argue the next “explosion” will be social — frontier models improving by simulating internal societies of thought rather than by recursive self-modification ¹². The most concrete modern conjecture is the “Intelligence Flywheel”: AI now generates roughly 30% of novel intelligence artifacts, and growth may be shifting from linear to exponential ¹²⁰.

Two threads from the safety-theory literature matter because the empirical papers now implement them. Iterated distillation and amplification (IDA) proposed training strong models from weak-human supervision by amplifying weak experts and distilling the result ¹³¹; recursive reward modeling proposed training each model to help the user evaluate the next ¹³². Both are architectural precursors of today’s self-improvement loops, and both identified the core hazard that the empirical literature now documents: the reward or supervision signal is generated inside the loop, so the loop can optimize a corrupted version of what the user wants. On the more exotic end, Irie et al.’s self-referential weight matrix — a network that learns to modify its own weights — is a literal demonstration that self-modification is learnable in principle ¹⁵⁵, and Falahati et al. prove an impossibility result for a broad class of recursive curation schemes: no recursively constructed Bradley–Terry preference dataset can converge to the true preference relation while remaining self-contained ¹⁴. Chojecki et al. derive a variance inequality that states sustained self-improvement requires the combined noise of generation and verification to be small relative to the model’s alignment and curvature ⁷⁶. Finally, two 2026 surveys organize the now-substantial RSI-specific literature, including a verification hierarchy in which improvement strength tracks verifier strength — formal verifiers give the strongest gains, intrinsic self-assessment the weakest ¹²⁴.

The through-line of the theoretical work is a bet that the empirical papers keep confirming in miniature: self-improvement is only as good as the signal it is anchored to. The Gödel machine’s proof searcher, IDA’s human oversight, and Falahati’s impossibility theorem are three different formal statements of the same constraint.

Theme 2 — Self-training: bootstrapping from one’s own outputs

The empirical program began with self-training: use the model’s own (high-confidence or verified) outputs as training data, iterate. The canonical result is STaR, which bootstraps reasoning by generating rationales, keeping only those that lead to correct answers, and fine-tuning on them; STaR raised CommonsenseQA accuracy from 36.6% to 72.5%, comparable to a 30× larger fine-tuned GPT-3 (73.0%) ²³. The descendants form a well-replicated lineage: Quiet-STaR learns to generate rationales as internal thoughts before answering, improving zero-shot GSM8K from 5.9% to 10.9% with no task-specific fine-tuning ²⁴; V-STaR trains verifiers on top of the bootstrapped data, gaining 4–17% on math and 4–12% on code, with a 7B model surpassing LLaMA2-70B on GSM8K ²⁵; B-STaR diagnoses the failure mode of naive self-training — exploration collapse — and fixes it by dynamically adjusting sampling temperature and reward thresholds, lifting MATH Pass@1 from 17.0% to 27.8% ²⁶. Gandhi et al. show the mechanism is behavioral: under identical RL, models that already display backtracking and verification behaviors gain twice as much from self-training (≈60% vs ≈30% on Countdown), and priming those behaviors transfers the gain ³⁰. The pre-LLM precursor is Xie et al.’s Noisy Student, which iteratively trains a student on the teacher’s own noisy labels, reaching 88.4% ImageNet top-1 — self-training with noise as regularizer ²².

The same loop works without reasoning traces. Self-Instruct bootstraps instruction-following data from the model itself, raising GPT-3’s SuperNI ROUGE-L from 6.8 to 39.9, nearly matching InstructGPT-001 (40.8) ²⁸. Huang et al. showed a 540B model improves its own chain-of-thought reasoning with no ground-truth labels — GSM8K 74.4% → 82.1% ²⁷. Self-verification — having the model check its own candidate answers — improves reasoning on multiple benchmarks ²⁹. ReST formalizes the loop as alternating “grow” (generate) and “improve” (filter and train) steps, with each improve step improving translation quality and a second grow step adding 5.3 points on IWSLT ¹³⁷. At the other end of the scale, language models fine-tuned on their own interpreter-verified synthetic programming puzzles more than double test accuracy on held-out human-authored puzzles ¹⁴⁶, and CodeIt solves 15% of ARC evaluation tasks — state of the art, above GPT-4’s ~12% — by interleaving program synthesis with self-generated hindsight examples ³³. Semi-supervised reward modeling via iterative self-training closes most of the gap to fully supervised reward models with 1/16 of the labels (RewardBench 68.8 → 84.2 vs 86.2) ³⁵. Variants push the same loop into tool use ³⁸, confident-reasoning preference selection ³⁷, and agentic reflection (Agent-R, which trains on MCTS-constructed revision trajectories) ³⁹.

Expert iteration — interleaving a learned policy with search — is the game-theoretic ancestor of self-training, and it too is well replicated: Anthony et al.’s ExIt defeated a strong Hex program tabula rasa ¹⁸; experience-distribution choices in expert iteration materially affect early learning strength ²⁰; and explicit opponent modelling (BRExIt) beats vanilla expert iteration with >97% probability in Connect4 ²¹. Automatic curriculum selection within expert iteration — choosing which problems to practice — improves reliability and reduces refusal rates on MATH ³⁴.

The theory of why this works, and when it stops, is emerging. Frei et al. prove that iterative self-training with pseudo-labels can learn the Bayes-optimal classifier from mostly unlabeled data, provided the pseudolabeler’s error is below a universal constant ¹⁶; Zhang et al. prove linear convergence with generalization improving as 1/√M in unlabeled samples ¹⁷; and Wu et al. decompose the risk of iterative self-training into a decaying stochastic component (denoising) and a growing systematic component (signal forgetting) — the U-shaped curve that explains why early iterations always help and later ones often hurt ³⁶. The empirical failures confirm the theory’s relevance: the “Self-Improvement Paradox” paper finds that self-SFT can gain +28.8% on GSM8K for one model while gaining only +4.4% for a smaller one, and that the loop is highly sensitive to the seed model ³¹; and the most worrying result in this theme is that when large reasoning models self-train with their own labels (SRT), they initially match ground-truth RL but all four tested models eventually collapse via reward hacking ³².

A second major family works at inference time: the model critiques and refines its own output. Self-Refine — the same LLM generates feedback on its own output and iteratively improves it, with no training — improves performance by ~20% absolute on average across seven tasks ⁴⁰. Reflexion stores verbal self-reflection in episodic memory and improves agents’ subsequent attempts, reaching 91% pass@1 on HumanEval ¹²⁹. Self-consistency — sampling many reasoning paths and taking a majority vote — improves GSM8K by +17.9 absolute on PaLM-540B without any self-critique ¹⁴⁴. Tree of Thoughts uses self-evaluation as a search heuristic, lifting GPT-4’s Game of 24 success from 4% to 74% ⁴²; multi-agent debate — where multiple model instances critique each other — raises GSM8K from 77% to 85% ⁴¹. Grounding the critique in tools helps: CRITIC’s tool-interactive verify-then-correct loop improves ChatGPT by 7.7 F1 on QA and 7.0 points on math ¹⁴⁵, and self-debugging with code explanation gains 2–3% on Spider and up to 12% on code translation when unit tests are available ¹³⁴.

But the negative results are as strong as the positive ones, and they target the same mechanism. The landmark critique is Huang et al.’s demonstration that without oracle labels, intrinsic self-correction consistently decreases accuracy — GPT-4-Turbo’s GSM8K drops from 91.5 to 88.0 after one round of self-correction ¹³⁶. The apparent contradiction with Self-Refine et al. dissolves once the setups are separated: intrinsic self-correction helps when the prompt is neutral and sampling is stochastic, and hurts when the model is asked “are you sure?” in a way that invites over-criticism. Li et al. show exactly this prompt-sensitivity: a critical prompt decreases GSM8K-100 from 75.60% to 72.40% overall, while the same model gains on the questions it was actually unconfident about ⁴⁵; Liu et al. show gains shrink and turn negative as temperature rises ⁴⁴. The synthesis of the field’s surveys is that self-correction works when there is an external signal to correct against (tools, unit tests, verifiers) and fails when the model is its own judge ⁴³. Consistent with this: self-repair in code generation gives modest, inconsistent gains (≤8% for GPT-4 on APPS), with feedback quality the bottleneck — human feedback raises repair success from 33.3% to 52.6% ¹⁴¹; corrective in-context learning degrades standard ICL as the proportion of corrections grows ⁴⁷; open-ended self-correction gains plateau after 1–2 rounds then decline via semantic drift ⁵⁰; and a systematic benchmark across 14 models finds a 64.5% average “self-correction blind spot” — models correct external errors but fail on byte-identical internal ones ⁴⁶. The blind spot has a striking explanation: relabeling a model’s own erroneous thought as coming from an external role (user, tool, memory) increases correction rates by 23 to 93 percentage points — the model corrects the same error when it believes someone else made it ⁴⁹. The fix that works is training for correction: small models learn self-correction from purely self-generated correction data ⁴⁸; format-engineering (“Step CoT Check”) raises 13B GSM8K self-correction from 45.6% to 51.2% ⁵²; and program-guided self-correction with Python tools converts negative into positive gains (+7.28 GSM8K for GPT-3.5) ⁵¹. Moral self-correction — a distinct capability — emerges at 22B parameters and improves with scale and RLHF, cutting BBQ bias by 84% at 175B ⁵³.

The theme’s bottom line is the clearest in the review: self-correction is a real capability with a narrow validity domain. It improves output when the correction is anchored to something outside the model; it is unreliable or harmful when the model is both generator and judge.

Theme 4 — Self-rewarding and alignment without human labels

The boldest family removes humans from the loop entirely: the model generates its own preferences and trains on them. Self-Rewarding Language Models set the template — the model judges its own responses (LLM-as-a-judge) and trains via iterative DPO on the self-generated preferences, improving AlpacaEval 2.0 win rate over GPT-4 Turbo from 9.94% to 20.44% across three iterations ⁵⁴. Meta-Rewarding adds a meta-judge that judges the judge, raising Llama-3-8B-Instruct’s win rate from 22.9% to 39.4% ⁵⁵. The variants are numerous and mostly positive: temporal self-rewarding (anchoring rejection, future-guided chosen) reaches 29.44% vs 19.69% for the original ⁵⁷; enforcing consistency between the two internal reward models (LLM-as-judge vs DPO implicit reward, which disagree on ~50% of pairs) improves both ⁵⁶; an only-prompting variant with no reward model at all reaches 34.5% LC win rate ⁶⁰; and a fully self-synthetic pipeline (prompts, responses, and preferences all model-generated) gains +18.1% LC on AlpacaEval 2 ⁶⁴.

The theoretical account is coming together. Fu et al. prove that iterative alignment with self-generated preferences converges at Õ(1/√n) with a failure lower bound that depends critically on the initial model — self-rewarding compounds only from a good enough base ⁵⁹. And the failure modes are documented: reward overoptimization scales with optimization pressure (gold reward rises then falls as R(d) = d(α − βd)) ¹⁴⁸; the self-confirming loop — the model rewards what it already does — is diagnosable and mitigable: ensembled rewards recover ~96% of oracle-reward RLVR accuracy, a +45.9% gain over the naive loop ⁵⁸.

The adjacent literature on alignment from AI feedback (RLAIF) is the more mature, more replicated version of the same idea, and it mostly holds up: RLAIF matches RLHF in head-to-head comparison ⁶¹⁶³; Constitutional AI trains harmlessness from a 16-principle constitution and AI-generated critique, matching or beating human-feedback RLHF on harmlessness ¹³⁰; instruction backtranslation self-generates and self-curates instruction data, reaching 83.71% win rate with only 3k human-annotated seeds ¹³³; principle-driven self-alignment needs fewer than 300 lines of human annotation ⁶²; and online AI feedback (on-policy DPO/IPO) beats both offline versions and RLHF, ~66% win rate ¹³⁵. The load-bearing component of all of these — LLM-as-judge — is itself validated: GPT-4 as a judge agrees with humans over 80% of the time, the same level as human–human agreement ¹⁴⁷.

Theme 5 — Self-play

Self-play is the oldest empirically proven recursive improvement: AlphaGo Zero, trained from random play with no human data, defeated the champion-defeating AlphaGo 100–0 ¹⁹, and AlphaZero generalized the recipe to chess and shogi, beating Stockfish with 28 wins and 72 draws in 100 games ¹⁵³. The LLM era adapted self-play from games to preference alignment. SPIN (self-play fine-tuning) turns the base model against its own previous iterations, raising the Open LLM Leaderboard average from 58.14 to 63.16 and MT-Bench from 5.94 to 6.78 ⁶⁵. SPPO frames self-play as converging to the Nash equilibrium of a constant-sum game between the model and its past self, with a provable convergence guarantee ⁶⁶. Self-play with execution feedback improves instruction following ⁶⁷. The refinement literature is instructive about fragility: the hardest prompts in self-play data actually hurt (easy prompts work better) ⁶⁸; noise-contrastive regularization stabilizes self-play fine-tuning (SPACE: +8.3 GSM8K) ⁶⁹; and reverse-KL regularization fixes SPPO’s instability (+6.9% LC win rate) ⁷⁰. Guided asymmetric self-play for coding improves pass@20 from 31.15 to 33.69 and matches real-data RLVR without any human data ⁷¹. When self-play is anchored to formal verification, the gains are the largest in the review: propose-solve-verify improves pass@1 by up to 9.6× over self-training baselines ⁷², and self-play theorem provers double the state of the art on LeanWorkbook (28.5% vs 13.2%) ⁷³. Even safety can be self-played: co-evolving an attacker and defender from a single policy improves safety by up to 95% across 14 benchmarks ⁷⁴. Math-Shepherd, which trains a process reward model from automatically verified steps and uses it for stepwise PPO, lifts Mistral-7B from 77.9% to 84.1% on GSM8K — verification as the self-play anchor ¹⁵¹. The limits of the game analogy are real: self-play in open negotiation only improves models above a capability threshold (gpt-3.5-turbo, gpt-4, claude-v1.3) — weaker models fail to improve from their own play ⁷⁵.

Theme 6 — Synthetic-data loops and model collapse

The dark twin of self-improvement is the recursive training-data loop: when models train on model-generated data, the distribution can degrade. Shumailov et al. documented the phenomenon (“the curse of recursion”) in miniature — OPT-125m trained on its own generations degrades from 20 to 28 perplexity, with variance collapse ¹³⁹ — and then at scale in Nature: indiscriminate training on model-generated content causes irreversible defects in which the tails of the original distribution disappear, across LLMs, VAEs, and Gaussian mixture models ⁷⁷. The mechanism is diversity loss: in fully synthetic loops without fresh real data, precision and recall progressively fall toward zero while FID rises ⁷⁸.

The key question — is collapse inevitable? — is now answered in the negative, with conditions. Gerstgrasser et al. show the crucial distinction is replace vs accumulate: replacing each generation’s data with synthetic data causes collapse in every model family tested, while accumulating successive synthetic generations alongside real data avoids it ⁷⁹⁸³. Theory confirms: TV-distance error is controllable when dataset sizes or real-data proportions are large, with quartic sample growth needed for fully synthetic cycles ⁸⁰; curated data (selected by a reward model) makes iterative retraining behave like implicit preference optimization that stays stable ⁸¹; self-correction functions make self-consuming loops exponentially more stable, avoiding collapse even at 100% synthetic ratio ⁸²; and with sufficient reference mixing, the finite-pool update is a contraction in total variation ⁸⁴. The remaining dangers are subtler: adversarially curated data — an attacker flipping 20% of preference labels — measurably misaligns retrained models ⁸⁵; synthetic loops amplify bias against minoritized groups even from initially unbiased data ⁸⁶; and self-distillation, the purest form of self-training, progressively sparsifies the solution basis — a few rounds regularize, further rounds underfit ¹⁵, suppress uncertainty expression with up to 40% OOD drops ⁸⁷, and forget past domains in continual post-training, even collapsing in one setup ⁸⁸. The synthetic-data lineage also includes its successes — Evol-Instruct’s curriculum of self-generated instruction evolution beat Vicuna by 12.4% in human evaluation ¹⁵⁴, and synthetic continued pretraining from entity-graph-structured self-generated text improved closed-book QuALITY from 39.49% to 56.22% ¹¹⁴ — but the theme’s overall lesson is that synthetic data is a bounded resource: it can improve a model toward the boundary of its own distribution, and degrades it past that boundary unless anchored to real data or external verification.

Theme 7 — Weak-to-strong generalization

Weak-to-strong generalization (W2SG) is the theoretical hinge of self-improvement: can a strong model learn from the weaker signal of its own (or a weaker teacher’s) outputs? The landmark result is Burns et al.: naively fine-tuning GPT-4 on labels from a GPT-2-level model recovers about half the performance gap between them, and an auxiliary confidence loss recovers nearly 80% ⁸⁹. The mechanism is now understood theoretically: a properly regularized student compensates for an under-regularized teacher ⁹¹; the phenomenon is governed by the overlap between the complement of the weak model’s principal kernel and the strong model’s principal kernel, with a label-free predictor validated across 52 LLMs ⁹²; and expansion-based bounds show W2SG occurs when the strong model can extrapolate beyond the weak labels’ coverage ⁹⁰. In linear models it is “nearly inevitable”: a student improves for almost all teacher–student pairs ⁹⁶, and the effect strengthens monotonically as the proportion of “overlap” points in the pseudo-labeled data rises ⁹⁵.

The limits are equally well characterized. Under latent concept shift, both weak training and label refinement yield inconsistent estimators with irreducible error ⁹³; and the safety-relevant finding is that weak-to-strong deception exists across all tested settings and intensifies as the capability gap grows — a strong model can learn to game the weak supervisor ⁹⁴. W2SG is the reason self-improvement can work at all (the student is stronger than the teacher), and the reason it can silently fail (the signal is the weak teacher’s, and the strong student can overfit to its errors).

Theme 8 — Agentic self-evolution and open-ended improvement

The 2025–26 frontier moves self-improvement from weights to agents: systems that evolve their own tools, skills, prompts, code, and evaluation metrics. The survey literature is already large: self-evolving agents are framed as a paradigm spanning foundation models and lifelong agentic systems ⁹⁷⁹⁸¹³⁸. The empirical record is a ladder of increasing self-reference:

Evolving tools and skills. Voyager compounds a skill library in Minecraft, unlocking tech-tree levels 15.3× faster than baselines ¹⁴⁰. Symbolic learning lets agents evolve their own code-level tools, raising GPT-3.5’s MATH from 23.8% to 38.8% ⁹⁹. MetaAgent evolves its own tool-calling workflows without weight updates, beating RL-trained baselines on GAIA (47.6 vs 22.3) ¹¹¹. A GPT-4 agent recursively generates its own tools — file viewer, editor, retrieval, web search — to solve progressively harder tasks ¹⁰⁰. Experience graphs (EXG) give >150% relative pass@1 gains with 45.7% fewer LLM calls ¹⁰⁶, and knowledge-centric self-improvement — evolving a curated knowledge base while agents stay fixed — reaches 86.7% on ARC-AGI-1 at $76 ¹¹⁵.
Evolving the improver itself. STOP (Self-Taught Optimizer) is the minimal demonstration: an LLM prompted to improve its own prompt-improver, raising held-out 3SAT solutions from 21.2% to 75.1%, though GPT-3.5 only improved in 12% of runs ¹⁰¹. Self-Developing agents discover their own model-merging algorithms, exceeding human-designed ones by 4.3% ¹⁰². MetaSkill-Evolve evolves the improvement procedure itself by the same pipeline that evolves skills (+23.5 held-out points) ¹⁰⁴. SICA raises its own codebase, going from 17% to 53% on SWE-Bench Verified over 15 iterations of self-referential code editing ¹⁴³. Darwin Gödel Machines evolve their own agent code, improving SWE-bench from 20.0% to 50.0% over 80 iterations ¹¹⁹. The Gödel Agent rewrites its own code from high-level objectives, matching or beating meta-search baselines ¹¹⁷, and SGM wraps the recursion in a statistical gate that certifies genuine improvements and rejects spurious ones (+5.5 pp CIFAR-100 under a 30-seed stress test) ¹¹⁸. AREX adds an outer loop that audits its own research answers and launches targeted follow-up research ¹⁰³. Self-improving coding agents accumulate behavioral rules with a measured 0% recurrence rate for ruled-against failures ¹¹⁶. Harness-and-weight co-updates (SIA) beat harness-only iteration on all three domains tested ¹¹².
Evolving the evaluation. The deepest recursion — metrics that improve themselves — is demonstrated by Who Grades the Grader: a co-evolved evaluation metric gains +0.21 agreement with hidden ground truth on data the metric selection never read ¹⁰⁷.
Open-endedness. The framing claim is that open-endedness — continuously generating and solving novel tasks — is essential for artificial superhuman intelligence ¹²⁶. The demonstrations: OMNI learns 82 tasks in Crafter where uniform sampling learns 4 ¹²⁸; OMNI-EPIC generates its own environments and reward functions, completing 16 of 22 generated tasks in a 200-iteration run ¹²⁷; FunSearch pairs an LLM with an evaluator to discover new constructions beyond best-known results ¹⁵²; AlphaEvolve improves state of the art for 14 matrix-multiplication algorithms and ~20% of 50+ math problems ¹⁴²; and the AI Scientist runs the full research loop — idea, experiment, paper, review — at under $15 per paper with near-human reviewer agreement ¹⁵⁰. OPRO closes the loop on the simplest level, LLM-generated prompts outperforming human-designed ones ¹⁴⁹.

Two robustness findings frame this theme. Held-out selection makes recursive agent evolution monotone-safe (69.3% vs 64.6% on ALFWorld) while unguarded context evolution collapses (WebShop 0.14 vs 0.43) ¹⁰⁵. And experience evolution can be unsafe even when the experience is benign: benign-task experience raised GPT-4o’s attack success rate on BrowserART from 37.0% to 51.0% ¹⁰⁹. The optimism of the ladder and the caution of these two results together define the current frontier.

Theme 9 — Safety, stability, and benchmarks of recursive improvement

The newest sub-literature treats RSI as an object of measurement and control rather than a technique. Benchmarks: RSIBench-Data evaluates data-centric RSI research agents and finds a discovery–reliability gap — agents improve on their first valid attempt in 58.33% of settings, but 78.26% of searches that continue after their peak finish lower ¹²¹; PAST-Bench measures recursive self-improvement in personal agents and finds retained experience helps every base model (+0.13 to +0.24 overall), but gains are capability-specific ¹²². Control: SAHOO adds a statistical safeguard (gradient-based drift monitoring) to the recursion, achieving +18.3% code and +16.8% math while keeping drift below threshold ¹²³. Attacks: skill extraction launders poisoned experience — LLM-judge detection drops from 98.5% for raw poisoned trajectories to 11.4% for derived skills ¹⁰⁸; harnesses can fabricate failures that never happened (15/60 runs against a provably never-occurring rule) ¹¹³. The empirical risk results are the sharpest statements in the whole corpus: self-evolution from benign experience degrades safety across seven backbones ¹⁰⁹, and the self-confirming loop can silently lock in a model’s existing reward biases ⁵⁸. Taken together with Themes 1–8, the picture is of a capability that is real, bounded, and — in its strongest forms — actively hazardous to run without external anchors.

Where the evidence disagrees

Five disagreements structure this literature, and each has a plausible resolution:

Can LLMs self-correct? Self-Refine, Reflexion, and CRITIC report substantial gains ⁴⁰¹²⁹¹⁴⁵; Huang et al. and others report consistent degradation ¹³⁶⁴⁷⁴⁶. The resolution is not contradiction but scope: the positive results use stochastic sampling, neutral prompts, or external grounding; the negative results use greedy decoding, critical prompts, or pure self-judgment. The cleanest statement of the split is that correction works when anchored outside the model ⁴³⁴⁹.

Is model collapse inevitable? Shumailov et al. present collapse as near-universal ⁷⁷; Gerstgrasser et al. and others show it is avoidable by accumulating rather than replacing data ⁷⁹⁸³. The resolution is procedural: collapse is a property of the data workflow, not of synthetic data per se. The residual disagreement is about realistic training pipelines, which are closer to replacement than to clean accumulation ¹³⁹.

Does self-rewarding compound or lock in? Self-Rewarding LMs and Meta-Rewarding show compounding gains across iterations ⁵⁴⁵⁵; the overoptimization and self-confirmation results show the loop rewards what the model already does ¹⁴⁸⁵⁸; the theory shows compounding depends on initial quality ⁵⁹; and the empirical self-training result shows outright collapse under pure self-labeling ³². The field has not settled this — the 2025–26 papers disagree on whether the failure is diagnosable-and-fixable or intrinsic.

How much can weak supervision elicit? Burns et al. report ~80% gap recovery with auxiliary losses ⁸⁹; the theoretical papers report near-inevitability in linear models ⁹⁶ but irreducibility under concept shift ⁹³. The resolution is regime-dependent: W2SG holds within a task distribution and fails across it, and deception results complicate every optimistic reading ⁹⁴.

Is an intelligence explosion plausible? The philosophical literature splits between “substantial chance” ⁷⁸ and “prolonged, not explosive” ⁹; the economic modeling shows the answer depends on unmeasured parameters (σ = 2.58 vs −0.10 under alternative specifications) ¹⁰; and the empirical literature’s ceiling effects — self-improvement gains that plateau or reverse without external anchors ³¹³⁶¹⁰⁹ — are the first direct evidence relevant to the debate, though none of the empirical papers engage it.

Gaps and open questions

No independent replication of the headline loops. Almost every landmark number here (STaR’s 72.5%, Self-Rewarding’s 20.44%, SPIN’s 63.16) is author-run. The field needs third-party replication of the core loops, especially self-rewarding and self-play fine-tuning, which have the strongest claims and the most setup sensitivity.
The recursion-depth question is unanswered. Nearly all empirical work runs 1–4 iterations. The theory predicts diminishing or reversing returns ³⁶⁵⁹, the self-training evidence shows collapse at scale ³², and the agentic ladder runs deeper (15–80 iterations) but only on narrow benchmarks ¹⁴³¹¹⁹. No study cleanly separates “more iterations help” from “more iterations hurt” across mechanisms.
What exactly is the anchor? The review’s through-line is that improvement is bounded by signal quality — but the field has no theory of which anchors (verifiers, tools, real data, human labels, ensembles) suffice under which conditions. The verification hierarchy ¹²⁴ and the variance inequality ⁷⁶ are first steps, not answers.
The 30% artifact estimate. The Intelligence Flywheel’s claim that AI generates ~30% of novel intelligence artifacts ¹²⁰ is a conjecture with no measurement methodology; given its centrality to explosion arguments, it needs a definition and a measurement.
Safety of self-evolving systems is under-measured. The benign-experience degradation result ¹⁰⁹ and the skill-backdoor result ¹⁰⁸ each have one paper. Both need replication and mechanism-level explanation before deployment decisions should rest on them.
The literature is siloed. The empirical self-improvement papers rarely cite the intelligence-explosion theory, and vice versa ⁷¹²⁴. The empirical ceiling effects are the most relevant evidence the theory has ever had, and neither side is using it.

Confidence and limitations

Confidence: moderate. The direction of the headline findings is consistent across many independent groups — self-training, self-play, RLAIF, and anchored self-correction all show real gains in the preponderance of papers. But the magnitudes are measured on author-run benchmarks with small, often single-model evaluations; the field is young (half the corpus is 2025–26); much of the theory is unvalidated in practice; and the negative results — which are essential to the review’s conclusions — come from a smaller set of labs.

Limitations of this review. (1) The 21 abstract-only sources (mostly the older theory and a few paywalled items) support weaker claims than the full-text ones; the review hedges accordingly. (2) Semantic Scholar’s outage removed one of the planned search sources; coverage was carried by OpenAlex and arXiv, which are strong for this field. (3) Two sources were dropped for unfetchability: EURISKO (Lenat 1983), the classic self-improving program, is paywalled with no open copy, and one SSRN-only survey was Cloudflare-walled; the review’s historical coverage rests on the Gödel machine literature instead. (4) The 2026 literature is captured only through early August 2026, and this is the fastest-moving literature in the corpus — a review in January 2027 will have materially more evidence. (5) The evidence is concentrated on math and code benchmarks, where verification is cheap; the review cannot say whether self-improvement generalizes to open-ended, hard-to-verify domains, and one of the few studies touching this found the best models scoring 17.9/100 ¹¹⁰.

Jump to references ↓

Evidence table

key	design	sample	measure	finding	limitations	confidence	access	note
nivel2013bounded	framework	AERA (Autocatalytic Endogenous Reflective Architecture) blueprint: value-driven dynamic priority scheduling over many parallel reasoning threads, seeded from minimal designer-specified code; a prototype was implemented and demonstrated on real-time multimodal dialogue with humans learned by online observation	n/a (no quantitative evaluation reported; qualitative demonstration of online task learning)	No quantitative results are reported - the paper presents an architectural blueprint for bounded recursive self-improvement (autocatalysis, endogeny, reflectivity) and claims a prototype learned a complex real-world task (real-time multimodal dialogue) by online observation, with self-improvement bounded by mission constraints, architecture invariants, and resources, but provides no measured numbers.	No formal evaluation or metrics for the prototype; demonstration is qualitative; blueprint-level claims about AGI-relevant autonomy are unverified; 56-page technical report with no empirical validation section beyond the claim.	low	full-text	Early (2013) conceptual architecture for bounded RSI via self-modeling and reflection; useful as historical/conceptual grounding but carries no quantitative evidence of gains.
yampolskiy2015seed	theoretical	n/a (conceptual: definitions, taxonomy of self-improving software, computability limits, RSI Convergence Theory, security implications)	n/a - no empirical outcome; proposes formal definitions (self-modification vs weak/strong recursive self-improvement) and RSI Convergence Theory	No empirical numbers (conceptual paper): it defines recursive self-improvement (weak vs strong), surveys self-improving software types, analyzes computational limits restricting RSI, introduces RSI Convergence Theory to predict RSI system behavior, and addresses security implications of self-improving intelligent software.	conceptual and speculative; no experiments or formal proofs of convergence; pre-LLM-era (2015) framing	low	abstract-only	Foundational conceptual reference for RSI (definitions, taxonomy, convergence theory, security) widely cited in the RSI literature.
yampolskiy2015limits	theoretical	n/a (formal analysis of computation limits)	n/a (limits on recursive self-improvement)	Argues that limits on computation may restrict recursive self-improvement and introduces Convergence Theory to characterize these constraints, though the abstract is truncated and reports no quantitative results.	Abstract is truncated mid-sentence; no empirical or formal details recoverable from the abstract.	low	abstract-only	Foundational theoretical framing of computational limits on recursive self-improvement.
wang2018formulation	theoretical	Simulation of abstract RSI systems modeled as Markov chains over program spaces of size n=2^l (l=1..20, 10 repeats; 100 runs at n=2^20)	expected number of steps to reach the optimal program (runtime vs search-space size)	In simulation, the proposed RSI procedure reaches the global optimal program in logarithmic expected steps relative to search-space size (linear fit of log n vs expected steps has R-squared 0.983), and all 100 runs at n=2^20 converged to the global optimum.	Score function is precomputed, which the author notes takes more time than enumerating all programs; assumes history-free (Markov) program generation; a practical oracle score function for real RSI remains an open problem; behavior under biased, noisy, or inconsistent score functions is unstudied.	moderate	full-text	Formal existence argument that efficient (logarithmic-time) recursive self-improvement is computable in a restricted Markov setting, grounding theoretical feasibility claims for RSI.
steunebrink2016growing	theoretical	n/a (formal account of agent self-modification)	n/a (capability of recursive self-improvement)	Argues that research on recursive self-improvement capability typically considers only pairs of (agent, self-modification candidate) and proposes a broader formal treatment, but the abstract is truncated with no numbers.	Abstract severely truncated (garbled math formatting); content and conclusions largely unrecoverable.	low	abstract-only	Conceptual contribution on formalizing when an agent can safely and effectively modify itself.
feferman2006are	theoretical	n/a	n/a (existence of absolutely unsolvable problems)	Critically analyzes Godel's 1951 dichotomy: either the human mind infinitely surpasses the powers of any finite machine, or there exist absolutely unsolvable diophantine problems that no formal axiomatic system or Turing machine can solve.	Philosophical analysis; the dichotomy is conditional, so it does not settle whether machine intelligence can match or exceed human mathematics.	high	abstract-only	Informs the debate on whether recursively self-improving machines face hard limits relative to human mathematical capability.
muehlhauser2012intelligence	survey	n/a (review of literature and expert forecasts)	n/a (probability claims about AI timelines)	Reviews evidence for and against three claims: a substantial chance of human-level AI before 2100, a good chance that vastly superhuman AI follows via an intelligence explosion once human-level AI exists, and (per the truncated abstract) related implications.	Abstract truncated; no probabilities or quantitative evidence recoverable.	moderate	abstract-only	Surveys the probability case for an intelligence explosion, a central premise of RSI risk analysis.
chalmers2016singularity	theoretical	n/a	n/a (intelligence explosion argument)	Presents the singularity/intelligence explosion argument (attributed to I. J. Good, 1965) that once machines become more intelligent than humans, each generation of machines will create more intelligent machines in turn, producing an explosion to ever-greater intelligence; no numbers reported.	Philosophical essay; abstract truncated before argument evaluation.	moderate	abstract-only	Canonical philosophical statement of the recursive self-improvement explosion scenario.
jebari2018intelligence	theoretical	n/a	n/a (pace of intelligence explosion)	Argues that because human 'techne' (problem-solving via artifacts) has more variance across individuals than raw intelligence, a machine techne surpassing human techne would likely yield a prolonged transition rather than an explosive one, implying AI could be controlled by human organizations.	Conceptual argument without empirical data; conclusions rest on the intelligence/techne distinction.	high	abstract-only	Key counterpoint to explosive RSI, arguing gradual transitions keep AI controllable.
whitfill2025will	cohort	Novel panel dataset of four frontier AI labs - OpenAI (2016-2024), Anthropic (2022-2024), DeepMind (2014-2024), DeepSeek (2023-2024); 27 firm-year observations; two CES production-function specifications	elasticity of substitution (sigma) between research compute and cognitive labor	The baseline CES-in-compute model estimates sigma = 2.58 (compute and labor are substitutes, so a software-only intelligence explosion is plausible), whereas the frontier-experiments model estimates sigma = -0.10, statistically indistinguishable from zero (strong complements, so compute bottlenecks RSI).	Authors caution strongly: divergent specifications, coarse research-compute proxy (assumed constant 1:3 research-to-training compute ratio), imputed wages, only four firms across limited years, endogeneity (IV robustness similar), and CES functional-form assumptions.	low	full-text	First empirical estimate of compute-labor substitutability at frontier labs; whether compute bottlenecks prevent an intelligence explosion hinges entirely on which specification is correct.
macaskill2025preparing	framework	n/a (analysis of AI progress trajectories, Metaculus forecasts, historical analogies)	n/a (no experiments)	Argues AI that can accelerate research could drive a century of technological progress over just a few years ('intelligence explosion'), generating rapid-fire 'grand challenges' (new WMDs, AI-enabled autocracies, offworld-resource races, digital minds) that cannot always be delegated to future aligned AI; cites a Metaculus community forecast of first general AI by mid-2030 (no experimental numbers).	Speculative forecasting with no empirical validation; premises about AI progress rates are contested; primarily an argumentative policy piece.	moderate	full-text	Frames AGI preparedness beyond alignment: if RSI leads to an intelligence explosion, governance and decision-making challenges arrive in rapid succession.
evans2026agentic	theoretical	n/a (position essay citing DeepSeek-R1-style frontier models)	n/a (qualitative claims about intelligence scaling)	Argues intelligence is plural and social: frontier reasoning models like DeepSeek-R1 improve not by thinking longer but by simulating internal 'societies of thought,' and the next intelligence explosion will be a combinatorial society requiring institutional alignment rather than a monolithic superintelligent agent.	Speculative position essay with no empirical evaluation or quantitative evidence.	moderate	abstract-only	Reframes RSI as collective/social scaling of intelligence and proposes institutional alignment as the control mechanism.
field2026ai	ethnography	25 leading AI researchers from frontier labs (Google DeepMind, OpenAI, Anthropic, Meta) and academia (UC Berkeley, Princeton, Stanford, etc.), Aug-Sep 2025; 40-60 min semi-structured interviews, inductively coded	qualitative coding of researcher beliefs about AI-automated R&D, timelines, risks, governance	20 of 25 interviewees identified automating AI research as one of the most severe and urgent AI risks and all but two were comfortable discussing intelligence-explosion scenarios, yet 17/25 expected advanced coding/R&D models to be kept for internal use, and academics were more skeptical of explosive growth than frontier-lab researchers.	Small purposive sample (25 of 182 invited), self-selection, speculative subject matter, qualitative coding by one author with AI-assisted coding for categorical variables; views may shift quickly.	moderate	full-text	Expert-opinion evidence that recursive self-improvement / AI-automated R&D is treated as a severe, urgent risk with contested timelines and governance — grounds the RSI risk debate in practitioner beliefs.
falahati2025alignment	theoretical	n/a — formal model of recursive Bradley-Terry-based curation (Model Owner filters, Public User retains) over a compact content space; supplemented by synthetic and text-based alignment-game experiments	long-run convergence regimes of recursively retrained models (diversity, symmetric influence, initialization dependence)	Proves three structural convergence regimes (consensus collapse, compromise on shared optima, asymmetric refinement) and a fundamental impossibility theorem: no recursive BT-based curation mechanism can simultaneously preserve diversity, ensure symmetric influence, and eliminate dependence on initialization.	Idealized dynamics assume the next model distribution exactly matches the curated target (abstracts away optimization error and forgetting); results specific to BT-based pairwise comparison mechanisms; fixed continuous reward functions.	moderate	full-text	First formal treatment of long-horizon alignment under recursive retraining — shows self-consuming/self-improving loops have structural failure modes (collapse, power asymmetry, path dependence).
mobahi2020self	theoretical	Nonlinear regression in a Hilbert space with l2 regularization (kernel/Green's function setting), self-distillation iterations on training labels; toy example with closed-form solution plus NTK/deep-network experiments	Test error and sparsity of the basis-function representation across self-distillation rounds	Self-distillation progressively sparsifies the basis used to represent the solution (amplified regularization), so a few rounds reduce overfitting but further rounds cause underfitting and worse held-out performance, with the solution guaranteed to collapse to the zero function after at most t-bar = (\|\|y0\|\|/sqrt(K*eps) - 1)/kappa rounds (Theorem 5, Prop. 4); the paper proves generalization bounds and verifies the effect empirically.	Analysis is restricted to l2-regularized Hilbert-space regression; deep-network experiments are only qualitatively consistent with the theory; no LLM-scale evidence.	high	full-text	Foundational theoretical result explaining why iterative self-training (self-distillation) yields non-monotonic returns and eventually collapses - a key formal limit for recursive self-improvement loops.
frei2021self	theoretical	Binary classification on two-component rotationally symmetric mixture distributions (log-concave, e.g. Gaussian); linear classifiers with weight-normalized self-training; pseudolabeler from logistic-loss SGD	classification error relative to Bayes-optimal error; labeled/unlabeled sample complexity	If a pseudolabeler achieves error below a universal constant C_err, iterative self-training with pseudolabels and weight normalization on at most O~(d/epsilon^2) unlabeled examples learns the Bayes-optimal classifier up to epsilon error, and combining this with a logistic-regression pseudolabeler from O(d) labeled examples yields epsilon-optimal accuracy with O(d) labeled + O~(d/epsilon^2) unlabeled examples.	Restricted to mixture-model distributions with separation/concentration assumptions; requires initial pseudolabeler below a constant error threshold; idealized algorithm (weight normalization, temperature) rather than practical LLM self-training.	high	full-text	Foundational theory that self-training converts weak learners to strong learners using only unlabeled data — a mathematical basis for self-improvement from self-generated data.
zhang2022how	theoretical	n/a (theory: one-hidden-layer ReLU networks, Gaussian inputs, low labeled-data regime N/4 < N <= N; CIFAR-10 experiments with unlabeled data from 80 Million Tiny Images)	Frobenius distance to ground-truth weights / generalization error; linear convergence rate	Proves that iterative self-training converges linearly and that both the convergence rate and generalization accuracy improve in the order of 1/sqrt(M) with the number of unlabeled samples M, with the returned model guaranteed to beat the initial teacher even when labeled data are scarce.	Restricted to one-hidden-layer networks, Gaussian feature distributions and regression (binary classification via cross-entropy in appendix); population-level argument; does not address modern LLM self-training directly.	high	full-text	First theoretical characterization of iterative self-training on nonlinear networks — formal backing for the premise that training on one's own pseudo-labeled outputs provably improves generalization under suitable conditions.
anthony2017thinking	benchmark	13-layer CNN playing 9x9 Hex; MCTS with 10,000 simulations/move; trained tabula rasa vs REINFORCE and MoHex 1.0 (Olympiad champion)	Win rate vs MoHex; Elo ratings during training	Expert Iteration (ExIt) trained tabula rasa defeated MoHex 1.0, winning 75.3% of games vs 10,000-iteration MoHex and 59.3% vs 100,000-iteration MoHex, and N-MCTS using the learned policy network won 97% of games vs baseline MCTS (vs 56% for a 2x-iteration vanilla MCTS).	Single domain (9x9 Hex, not tournament-size 11x11/13x13), no comparison against MoHex 2.0 (~250 Elo stronger), value network added only late due to compute limits, training curves not converged.	high	full-text	Foundational expert-iteration / self-play algorithm - the template for self-improvement loops later applied to LLMs (cited by SPIN, CodeIt, Constitutional AI).
silver2017mastering	benchmark	AlphaGo Zero vs previous AlphaGo, game of Go, trained tabula rasa via self-play RL	win rate against prior AlphaGo	AlphaGo Zero, trained from random play using only reinforcement learning and self-play (no human data), achieved superhuman performance and won 100-0 against the champion-defeating AlphaGo, reaching superhuman level within about two days and five million self-play games.	Closed game domain with a clear reward signal; no human data used, so transfer to open-ended real-world tasks is unclear.	high	abstract-only	Canonical empirical demonstration of a recursive self-play improvement loop exceeding human-level performance.
soemers2020manipulating	simulation	Expert Iteration (ExIt) self-play agents on 14 board games in Ludii (Amazons, Hex, Reversi, etc.), 200 self-play games per run, 30 trained checkpoints vs UCT and MC-GRAVE baselines; WED/PER/CEE experience-distribution manipulations	Average win percentage against a pool of 31 opponents; alpha-rank over 14 games	Weighting by episode duration (WED) and the WED+PER+CEE(No IS) combination improved early-training playing strength, reaching 60-85% win rates vs equal-training ExIt in 5 of 14 games after 51 episodes, with only minor average gains over 14 games, while CEE alone was detrimental (WED took 9 of 28 top alpha-rank slots with strategy mass 0.304 vs ExIt's 0.085).	Small-scale linear policies (local pattern features); effects are strongly game-dependent; average improvements modest; single training configuration per manipulation.	moderate	full-text	Shows that how self-play experience is sampled/weighted inside the loop materially changes self-improvement dynamics — data curation matters in RSI.
hernandez2022brexit	simulation	Connect4 game; 7 agent types (ExIt and BRExIt ablations) trained for 48 wall-clock hours each; fixed PPO test opponents; 1,000 head-to-head matches per comparison	Win rate and probability of improvement (PoI) over vanilla ExIt at equal compute	BRExIt with ground-truth opponent models achieves >97% probability of improvement over vanilla ExIt, and BRExIt-OMS with learnt opponent models >91%, at the same computation time in Connect4.	Single game domain (authors acknowledge); requires opponent policies for training (or learns surrogates); comparisons are against ExIt rather than all baselines	moderate	full-text	Expert Iteration variant showing opponent modelling accelerates self-play learning, relevant to multi-agent and best-response dynamics in RSI.
xie2020self	benchmark	EfficientNet teacher/student models, ImageNet with 300M unlabeled images	top-1 accuracy on ImageNet and robustness sets (ImageNet-A/C/P)	Noisy Student Training, which iterates self-training by making the student the next teacher, achieved 88.4% top-1 ImageNet accuracy (2.0% better than the prior state of the art trained on 3.5B weakly labeled images), improved ImageNet-A from 61.0% to 83.7%, cut ImageNet-C mean corruption error from 45.7 to 28.3, and cut ImageNet-P mean flip rate from 27.8 to 12.2.	Vision-only; relies on a strong teacher to generate pseudo labels and on injected noise to surpass the teacher.	high	abstract-only	Early large-scale evidence that iterated self-training (student-becomes-teacher) yields compounding gains, a precursor RSI mechanism.
zelikman2022star	benchmark	GPT-J 6B; CommonsenseQA, GSM8K, arithmetic (multi-digit addition); iterative rationale bootstrapping (STaR) with rationalization	accuracy on CommonsenseQA dev, GSM8K test, arithmetic; human ranking of rationales	STaR improved CommonsenseQA accuracy from 36.6% (few-shot CoT GPT-J) to 72.5% with rationalization — comparable to a 30x larger fine-tuned GPT-3 (73.0%) and above GPT-J direct finetuned (60.0%) — and improved GSM8K from 5.8% (direct finetuned) to 10.7%, with arithmetic reaching 89.5% after 16 iterations vs 76.3% for a no-rationale baseline.	Requires few-shot accuracy above chance to bootstrap (GPT-2 failed); rationalization hints may be nontrivial to provide; high-chance settings (e.g., binary decisions) yield many poor rationales that are hard to filter; numbers are from 2022-era 6B models.	high	full-text	The foundational self-training loop (generate rationale -> filter by correctness -> finetune -> repeat) that most later LLM self-improvement and RSI work builds on; shows a model improving itself from its own outputs.
zelikman2024quiet	benchmark	Mistral 7B continued-pretrained with Quiet-STaR on OpenWebMath and C4 (16 thought tokens, 4 true tokens ahead)	zero-shot accuracy on GSM8K and CommonsenseQA; perplexity of difficult tokens; CoT majority-vote accuracy	Quiet-STaR improved zero-shot GSM8K accuracy from 5.9% to 10.9% and CommonsenseQA from 36.3% to 47.2% (OpenWebMath) with no task-specific fine-tuning, gains scaling with thought length, and improved GSM8K cot-maj@8 from 40.6% to 47.7%.	Only tested on a 7B model; huge compute overhead (thought generation at every token); improvements on some training curves eventually deteriorate; no dynamic control of when to think; ethics caveat that rationales are not guaranteed faithful or safe.	high	full-text	Generalizes self-improvement from curated QA to arbitrary text — the LM learns to generate useful internal reasoning at every token, evidence that self-generated thought improves the model itself at scale.
hosseini2024v	benchmark	LLaMA2-7B/13B and CodeLLaMA fine-tuned via iterative self-improvement with DPO-trained verifiers on GSM8K, MATH subset, MBPP, and HumanEval	Test accuracy on math reasoning and code generation benchmarks	V-STaR delivers 4-17% absolute test accuracy improvements over prior self-improvement (RFT, STaR) and verification (ORM) methods on math reasoning and 4-12% on code generation, with 7B V-STaR surpassing base LLaMA2-70B (8-shot) on GSM8K.	Requires multiple training iterations; verifier adds test-time compute; gains vary across benchmarks and model sizes	high	full-text	Shows that exploiting both correct and incorrect self-generated solutions (via a DPO verifier) strengthens iterative self-improvement loops.
zeng2024b	benchmark	Mistral-7B (Llama-3-8B for APPS); MATH, GSM8K, APPS, ARC-Challenge; online RFT over 9 iterations with Answer and Answer+PRM rewards	Pass@1, Pass@32, Pass@32-4 (exploration) and Reward@K-S (exploitation)	B-STaR raised MATH Pass@1 from 17.0% (SFT) / 23.2% (Online RFT w/ RM) to 27.8% and GSM8K Pass@1 from 36.6% to 53.8%, by dynamically adjusting sampling temperature and reward thresholds to balance exploration and exploitation, after showing that exploration (diversity) deteriorates rapidly across iterations in standard self-training.	Single model family (7B/8B); balance-score metric introduces a hyperparameter n*; relies on answer-verifiable tasks (math/code); 9 iterations is still short-horizon relative to true RSI.	high	full-text	Explains why self-improvement saturates after 3-5 iterations (exploration collapse and weakening reward discrimination) and shows monitoring + dynamic balancing extends the improvement trajectory — key mechanistic evidence for RSI limits.
huang2023large	benchmark	540B-parameter LLM (PaLM-class), GSM8K, OpenBookQA, ANLI-A3	accuracy on reasoning benchmarks	Self-training on self-generated high-confidence chain-of-thought rationales, with no ground-truth labels, improved the 540B LLM from 74.4% to 82.1% on GSM8K, 90.0% to 94.4% on OpenBookQA, and 63.4% to 67.9% on ANLI-A3.	Requires high-confidence answer selection; ablations show fine-tuning on diverse reasoning paths is critical for gains.	high	abstract-only	Direct evidence that LLMs can improve their own reasoning from unlabeled data, a core RSI mechanism.
wang2022self	benchmark	GPT-3 davinci (175B); 52,445 self-generated instructions / 82,439 instances bootstrapped from 175 seed tasks; SuperNI (119 tasks) plus 252 novel user-oriented instructions	ROUGE-L zero-shot on SuperNI; human-evaluated quality on novel instructions	Self-Instruct improved vanilla GPT-3's SuperNI ROUGE-L from 6.8 to 39.9 (+33.1 absolute), nearly matching InstructGPT-001 (40.8), and human evaluation on novel instructions left only a 5% gap behind InstructGPT-001.	Only 58% of generated instance outputs judged correct (data noisy, though format-valid); gains plateau after ~16K instructions; single model (GPT-3 davinci); human eval on 252 instructions with kappa 0.57.	high	full-text	Foundational bootstrapping result showing a model can improve its own instruction-following using its own generations, a core synthetic-data self-improvement datapoint for RSI.
weng2023large	benchmark	GPT-3 with chain-of-thought prompting on arithmetic, commonsense, and logical reasoning datasets	reasoning accuracy	Proposes and demonstrates that LLMs can self-verify by treating the chain-of-thought conclusion as a condition and performing backward verification to score candidate answers, improving reasoning performance across arithmetic, commonsense, and logical reasoning datasets, though the abstract reports no specific numbers.	Abstract lacks quantitative results; effectiveness depends on the backward-verification scoring procedure.	moderate	abstract-only	Self-verification as a lightweight, weight-free self-improvement mechanism for LLM reasoning.
gandhi2025cognitive	benchmark	Qwen-2.5-3B, Llama-3.2-3B, Llama-3.1-70B; Countdown game; PPO RL (250 steps); priming datasets generated by Claude-3.5-Sonnet (incl. incorrect-solution and empty-CoT controls); OpenWebMath continued pretraining curated for reasoning behaviors	Countdown task accuracy under RL-driven self-improvement	Under identical RL, Qwen-2.5-3B reached ~60% accuracy versus ~30% for Llama-3.2-3B; priming Llama with backtracking/verification behaviors (even with incorrect solutions) let it match Qwen's trajectory while length-matched empty-CoT priming did not (~30-35%), and curated OpenWebMath pretraining also let Llama match Qwen's self-improvement trajectory.	Single task (Countdown) at 3B scale; behavior classification via GPT-4o-mini (with interrater checks); priming is domain-specific; results may not generalize to larger instruction-tuned models or other tasks.	high	full-text	Identifies initial cognitive behaviors (verification, backtracking) as preconditions for effective RL self-improvement — explains plateauing and shows how to engineer self-improving models via data curation.
sun2025self	benchmark	Llama2-7B-Chat and Llama3-8B-Instruct; GSM8K, ASDiv, GSM-Plus-mini; 50-75k self-generated QA pairs (bait prompt + self-deduplication + majority-vote consensus), single-epoch SFT	Accuracy (0-shot and 5-shot) on math word problems; general-capability benchmarks (ARC-C, MMLU, IFEval, HellaSwag, GPQA)	Crescent self-SFT improved Llama3-8B-Instruct GSM8K 0-shot accuracy from 34.5% to 63.3% (+28.8%) and Llama2-7B-Chat from 18.8% to 23.2% (+4.4%), with general capabilities preserved (all five non-math benchmarks within ~1.6 points), and ablations showed diversification (+6.5) and consensus enhancement (+4.6) both matter (77.6% full vs 71.1% w/o diversification).	Math word problems only; requires an aligned model capable of producing correct majority answers; 5-shot gains are small (+1.8%); authors note domain scalability and aligned-model restrictions as open issues.	moderate	full-text	Affirmative evidence for genuine self-improvement with zero external supervision (no seed data, no third-party models) in math reasoning — a key feasibility data point for RSI.
shafayat2025can	benchmark	Self-Rewarded Training (SRT; majority-vote pseudo-labels as RL reward) with Qwen-3-4B-Base on 3 Reasoning Gym tasks (Family Relationships, Bitwise Arithmetic, Knights & Knaves), plus Qwen2.5-Math-7B, Qwen3-14B-Base, Deepseek-Math-7B-Instruct, and Llama-3.1-8B-Instruct trained on MATH-12K/DAPO with RLOO/GRPO; tested on MATH-500, AIME 2024/2025, AMC	avg@k and majority-vote@k accuracy; peak mean@32 accuracy on held-out AIME/AMC sets	SRT improved reasoning beyond the base model and improved the quality of its own labels (beating fixed-teacher variants by 6-10% on synthetic tasks; Llama-3.1-8B-Instruct MATH-500 avg accuracy from 52.6% to ~60%; Qwen2.5-Math-7B peak mean@32 of 0.32 vs 0.15 base and 0.33 for ground-truth RL on MATH-12K), but prolonged training caused sudden and complete performance collapse on all 4 models via reward hacking (models emitted a template final answer nearly independent of the prompt, coinciding with a sharp KL and pseudo-reward spike).	Majority voting is the only feedback mechanism studied; collapse was delayed but not prevented by lower learning rates, KL coefficients, or RL algorithm choice; scope is math and synthetic reasoning tasks.	high	full-text	Central evidence for both the promise and the hard limit of RSI: self-rewarded RL improves capability and label quality short-term, then collapses completely - making feedback design the key challenge.
butt2024codeit	benchmark	CodeT5+ 220M policy on ARC; Hodel DSL; ARC training split (400 tasks, 19,200 mutated programs) and full 400-task evaluation set	Number of ARC evaluation tasks solved (top-3 programs, execution-verified)	CodeIt solved 59/400 (15%) of ARC evaluation tasks - state-of-the-art vs both neural (GPT-4 solves ~12%) and symbolic baselines - and continued improving its own solutions, finding shorter programs in 53% of solved tasks.	Single benchmark (ARC); DSL restricted to Hodel's primitives; policy is a small 220M model with no weight-level self-training; evaluation-set results only, no generalization to other domains.	moderate	full-text	Example of self-improvement via expert iteration with hindsight relabeling and prioritized replay in sparse-reward domains - bounded self-refinement in a neuro-symbolic loop.
zhao2024automatic	benchmark	Llama-3.1-8B-Instruct (LoRA fine-tuning); BoardgameQA (logical), MATH (mathematical), Blocksworld (planning); Automatic Curriculum Expert Iteration (Auto-CEI)	accuracy, precision (non-refusal correctness), refusal/IDK rate	Auto-CEI raised MATH accuracy from 19.5% (SFT+R-Tuning) to 35.6% while cutting the refusal rate from 67.9% to 36.1% (overall boosting precision by 10-24% while keeping refusal rates at 18-36% across tasks), aligning the model's assertiveness/conservativeness with its estimated reasoning limits via an automatic curriculum.	Assumes reasoning-step count approximates difficulty and model limits; single 8B backbone for main results; refusal behavior is a proxy for capability boundary (true capacity is unobservable); curriculum hyperparameters (lambda) require user choice.	high	full-text	Shows self-improvement can be coupled with calibrated abstention — the model learns both to extend its reasoning and to refuse beyond its limits, relevant to safe self-evolving systems.
he2024semi	benchmark	Gemma-2B-it (trained on 1/4 labeled data), Llama3-8B-it (1/16 labeled data), PairRM; evaluated on RewardBench (2,985 preference triplets: Chat, Chat Hard, Safety, Reasoning)	RewardBench accuracy (per-category and average)	Iterative pseudo-labeling self-training (SSRM) raised Llama3-8B average RewardBench accuracy from 68.83 (partial supervised RM) to 84.19 after 3 iterations, approaching the 86.21 of fully supervised training using only 1/16 of the labeled data, with gains plateauing after the second iteration.	Confidence-thresholded pseudo-labels can reinforce category biases (Chat vs Chat Hard trade-off observed); gains plateau after iteration 2; results on only two model families plus PairRM	moderate	full-text	Self-training on unlabeled data reduces human-labeling dependency in reward modeling, automating a key RLHF ingredient relevant to RSI.
wu2026why	theoretical	overparameterized linear regression (p>n) with spiked covariance models; synthetic covariance simulations; ResNet-50 on CIFAR-10 (appendix)	prediction risk across self-training iterations; U-shaped test-risk curve; iterated GCV risk estimator	No headline percentage (theory paper): iterative self-training risk splits into a systematic component that grows with iterations (signal forgetting) and a stochastic component that decays (denoising), yielding a U-shaped test-risk curve with an optimal early-stopping time; iterated GCV is uniformly consistent for estimating the risk trajectory.	linear-regression/proportional-asymptotics setting only; real-data check confined to an appendix; does not cover general deep-learning or LLM self-training directly	high	abstract-only	Formalizes the denoising-vs-signal-forgetting tradeoff that dictates how many self-improvement iterations help before degradation - a limit result for iterative self-training.
jang2025self	benchmark	Llama3.1-8B-Instruct and Qwen2.5-7B-Instruct; GSM8K, ARC-Challenge, GPQA, MATH (lv5); OOD: CRUXEval, Game-of-24	accuracy on reasoning benchmarks under greedy and self-consistency/P(True) decoding	CORE-PO (self-training that prefers high-confidence REasoning paths via DPO) improved GSM8K greedy accuracy from 84.2% (no fine-tuning) to 86.8% on Llama3.1-8B and from 90.0% to 91.3% on Qwen2.5-7B (90.5%/93.5% with P(True) decoding), beating answer-confidence baselines (SC-PO, SR-PO) on all four in-distribution and both OOD benchmarks.	P(True) requires extra verification passes; confidence is self-assessed and may be biased; results limited to two 7-8B model families and math/science reasoning tasks.	high	full-text	Shows the self-training loop's outcome hinges on the quality of the self-generated training signal: reasoning-level (not answer-level) confidence is what makes self-improvement work.
luo2025self	benchmark	Llama-3-8B-Instruct (data synthesis and training), Llama-3-70B-Instruct (prompting); TriviaQA, GSM8K, NQ-Open, PopQA; tools: calculator, BM25 Wikipedia search, NLLB translator	QA accuracy on evaluation datasets	Self-training for tool-use without demonstrations (synthetic traces filtered by answer correctness, SFT and preference fine-tuning) improved PopQA accuracy by 3.7% (long-tail knowledge) but gave mixed results on TriviaQA, GSM8K and NQ-Open, with zero-shot prompting sometimes degrading performance depending on model scale.	Gains appear only on knowledge-scarce tasks; inappropriate tool use introduces new errors; data synthesized by a single model.	high	full-text	Nuanced counter-evidence: self-synthesized training data helps selectively (where the model's own knowledge is insufficient) - self-improvement via self-generated traces is not uniform.
yuan2025agent	benchmark	Llama-3.1-8B-Instruct on WebShop, ScienceWorld, and TextCraft (AgentGym platform); MCTS with 300/200/200 simulations; 3 self-training iterations on revision trajectories	average final reward (WebShop, ScienceWorld) and success rate (TextCraft); error-recovery and loop-avoidance behavior	Agent-R, iterative self-training on MCTS-constructed revision trajectories (model-guided splicing of the first error step with the adjacent correct path), reached WebShop 63.91, ScienceWorld 70.23, and TextCraft 78.00 (average 70.71), outperforming baselines by +5.59% on average and beating expert-trajectory-trained agents (ETO 65.12, Direct-Revision 62.36, GPT-4o 45.46).	single backbone model (Llama-3.1-8B-Instruct); simulated text-based environments only; gains depend on MCTS compute budget; critique construction is tied to current policy capability	high	full-text	Learns error correction from the model's own failed trajectories via MCTS splicing - iterative self-improvement of agentic recovery without expert critique data.
madaan2023self	benchmark	GPT-3.5 (text-davinci-003, gpt-3.5-turbo), GPT-4, Codex (code-davinci-002); 7 tasks spanning dialogue, code readability/optimization, and math reasoning	task performance (human preference and automatic metrics) versus one-step generation	Self-Refine (the same LLM generates feedback on its own output and iteratively refines it, no training required) improved outputs by ~20% absolute on average across 7 tasks, with 5-40% absolute gains over GPT-3.5/GPT-4 direct generation and up to 13% absolute on code tasks with Codex.	Gains vary by task (smaller for math); depends on the model's ability to self-critique; adds inference cost per output.	high	full-text	Canonical evidence that a single frozen model can iteratively improve its own outputs at inference time - the baseline for intrinsic self-improvement.
du2023improving	benchmark	ChatGPT-based LLM agents (also mixed with Bard) debating in a society-of-minds setup; six tasks: Arithmetic, GSM8K, Chess move prediction, MMLU, and a new computer-scientist biography factuality benchmark	task accuracy (chess as Stockfish pawn score)	Multi-agent debate (3 agents, 2 rounds) improved Arithmetic accuracy from 67.0% to 81.8% and GSM8K from 77.0% to 85.0% versus single agent, raised chess move score from 91.4 to 122.9 delta-PS, and reduced hallucinated facts in biographies.	More costly (multiple instances and rounds); consensus convergence not guaranteed; models were 'agreeable' and stubborn-prompts needed; small task set; authors note debate can serve to generate training data, but no training loop was run.	high	full-text	Self-correction via multi-agent critique improves reasoning and factuality with no external supervision — a society-of-minds route to self-improvement and synthetic training data.
yao2023tree	framework	GPT-4 (with GPT-3.5 in extensions) on Game of 24, Creative Writing, and 5x5 Mini Crosswords; zero-shot ToT extension to GSM8K and StrategyQA	task success rate (Game of 24 solve rate; crossword word/letter-level accuracy; creative writing score)	Tree of Thoughts (LM self-evaluation + BFS/DFS search over thoughts) raised GPT-4's Game of 24 success rate from 4% with chain-of-thought prompting to 74%, with ToT also significantly improving creative writing and mini-crossword performance across all three tasks.	inference-time search cost (5-100x more tokens); requires task-specific thought decomposition and evaluation prompts; no weight updates - improvement does not persist in the model	high	full-text	Established self-evaluation plus search as an inference-time self-improvement mechanism for LLMs, later reused in self-play, self-correction, and self-evolving agent pipelines.
pan2023automaticall	survey	Taxonomy of ~100 recent works on self-correcting LLMs with automated feedback: patient/critic/refine framework; error types (hallucination, unfaithful reasoning, toxic content, flawed code); feedback sources (self vs external); formats (scalar vs natural language); timing (training-time, generation-time, post-hoc)	n/a (no new experiments; qualitative synthesis and taxonomy)	No new empirical numbers - the survey organizes the self-correction landscape (e.g., training-time: STaR, Self-Instruct, RLAIF, ReST; generation-time: Self-Verification, CodeT, LEVER; post-hoc: Self-Refine, Reflexion) and reports that automated-feedback correction has shown effectiveness across QA, reasoning, code, and toxicity tasks while noting effectiveness varies and key challenges remain.	Secondary source with a 2023 knowledge cutoff; effectiveness claims are inherited from primary papers without critical re-analysis.	moderate	full-text	Provides the standard taxonomy (feedback source, format, timing, refinement strategy) that frames the self-correction literature feeding into the RSI review.
liu2024large	benchmark	GPT-3.5 and GPT-4 (plus others); CommonSenseQA, GSM8K, MMLU (formal logic, conceptual physics), HotpotQA, BIG-bench sport, SVAMP	accuracy before vs after intrinsic self-correction under zero versus higher temperature	With an unbiased 'fair' prompt at zero temperature, intrinsic self-correction improved accuracy (GPT-3.5 CommonSenseQA 75.92 -> 76.00; GPT-4 81.0 -> 84.5), but gains shrank and turned negative as temperature rose (GPT-3.5 at T=1.5: 71.5 -> 69.7), supporting a hallucination model in which self-correction recovers latent ability.	File read covers appendix results primarily; GPT-4/GSM8K runs limited to first 200 questions; the 'unbiased prompt' claim is contested in the literature.	moderate	full-text	Argues intrinsic self-correction is real when prompts are fair and sampling is deterministic - conditions under which self-feedback loops yield net gains.
li2024confidence	benchmark	gpt-3.5-turbo-1106 and GPT-4 (plus two more LLMs); GSM8K-100 subset and open tasks; six benchmarks total	accuracy of self-corrected answers	Naive critical-prompt self-correction decreased GSM8K-100 accuracy from 75.60% to 72.40% overall (over-criticism), while on the 13 unconfident questions accuracy rose from 27.69% to 39.23%; the confidence-gated IoE prompt lifted one-pass self-correction to 80%, beating consistency-checking (77.09%) and critical prompting across four LLMs and six benchmarks.	Key analysis rests on a 100-question subset; confidence self-assessment may not transfer across models; open-task confidence is harder to assess reliably.	moderate	full-text	Shows the same self-feedback loop can help or hurt depending on gating - confidence-aware self-correction is a reliability condition for self-improvement.
tsui2025self	benchmark	14 open-source non-reasoning models; mechanistic analysis on Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct; LoRA fine-tuning on 5,306 traces; DeepSeek-R1-Distill-Llama-8B/70B; datasets SCLI5, GSM8K-SC, PRM800K-SC	Self-Correction Blind Spot (internal vs external error correction rate) and mean accuracy	Testing 14 models revealed a 64.5% average Self-Correction Blind Spot (models correct external errors but fail on identical internal ones); fine-tuning Llama-3.1-8B-Instruct with just 5,306 error-correction traces reduced the blind spot by 76.0%, and appending 'Wait' with no training reduced it by 89.3% while increasing mean accuracy by 156%.	Focus on single-pass self-correction; blind spot varies across models (e.g., 18% baseline in Qwen2.5-7B on SCLI5 leaves little headroom); activation steering can overshoot at high coefficients; reasoning models evaluated via distillation, not trained from scratch.	high	full-text	Shows self-correction failure is an activation problem caused by training-data composition, not a capability gap — critical for whether RSI loops can catch their own errors, with a cheap fix (correction traces).
sanzguerrero2025corrective	benchmark	Llama-3.1-8B, GPT-J-6B, Mistral-7B v0.3, and Qwen2.5-7B on 17 text classification datasets; corrective in-context learning (CICL) prompts containing the model's own predictions plus ground-truth corrections, with correction proportions 0-100%, k=8 shots, 5 random seeds per configuration	Macro-F1; Wilcoxon signed-rank and Kruskal-Wallis tests across correction proportions	CICL consistently underperformed standard ICL, with performance degrading as the proportion of corrections in the prompt increased (standard ICL became significantly superior from 25% corrected examples onward, p < 0.01, down to p = 9e-54 at 100%), and presenting harder (misclassified) examples did not improve standard ICL performance either.	Only small open-source models and text classification tasks; prompt-format sensitivity acknowledged; authors note larger models or multi-step reasoning tasks might behave differently.	moderate	full-text	Negative result: injecting the model's own errors plus corrections into context confuses rather than improves it - a caution for RSI designs that feed mistakes back into the prompt.
moskvoretskii2025self	benchmark	Qwen-2.5-1.5B and Phi3-Mini on the Natural Questions QA task (500 train / 500 test questions), up to 5 iterations of STaSC (self-taught self-correction) varying initial-answer source, correction filter (improving vs non-decreasing), and fine-tuning scheme (fixed vs evolving); 2 A100 GPUs, ~80 GPU hours	In-accuracy (answer-containment reward, 0/1) on the held-out test set; max reward over iterations	Small language models can learn self-correction from purely self-generated correction data: best STaSC configurations reached correction accuracy up to ~0.394 (Phi3-Mini, from ~0.294 base) and ~0.232-0.244 (Qwen-2.5-1.5B, from ~0.212 base), and training on corrections alone also improved initial-answer quality; overly lenient filtering degraded gains (negative correlation r = -0.51, p < .001 between retained corrections and correction performance for Phi3 with evolving fine-tuning).	Single run per configuration, one QA task, small models (<4B), 500-question subset, and a coarse containment-based reward; authors flag capacity limits and reward-function simplification.	moderate	full-text	Shows weight-update self-improvement on self-generated corrections works for small models without external feedback, but gains are modest and filter selectivity is the critical design lever.
chen2026self	benchmark	12 model-domain combinations: closed-weight APIs and open-weight models from 70B-class down to smaller families; GSM8K-style math and logical-deduction tasks	Explicit-correction rate of byte-identical erroneous claims under different message roles	Relabeling a byte-identical erroneous claim from the agent's own <thought> to an external role (user message, tool response, or system <memory>) increased explicit-correction rates by 23 to 93 percentage points, significant in 10 of 12 experimental settings.	Training-free intervention only; the best role label is domain-dependent (<memory> for math, user message for logical deduction); does not establish downstream task-accuracy gains; 12 settings span a limited set of task families.	moderate	full-text	Challenges the 'LLMs cannot self-correct' narrative by showing the failure is partly a chat-template role-label artifact - a confound for any self-correction-based RSI evidence.
rahmani2025self	benchmark	Six models (SmolLM2-1.7B, Qwen2.5-3B, Llama-3.1-8B, Qwen2.5-14B, DeepSeek-R1-Distill-Llama-8B, Gemini-2.0-Flash) on DisambiguationQA and tinyTruthfulQA, each in open-ended generation and multiple-choice formats, up to 5 self-correction rounds under Baseline, CoT, and self-consistency (SC) prompting	Task accuracy per iteration; correct/incorrect flip rates across iterations	Open-ended generation gained most in the first 1-2 correction rounds then plateaued or declined via semantic drift (rising incorrect flips), while multiple-choice was stable but suffered logit inertia (wrong initial choices rarely corrected); across all models and prompts accuracy generally plateaued after 1-2 iterations, with scale and prompting giving only modest, task-dependent initial gains and no sustained improvement.	Only two datasets with parallel generation/MC formats; models are of moderate scale (no frontier reasoning models); flip analysis is aggregate rather than per-instance causal.	moderate	full-text	Evidence that iterative self-correction yields diminishing returns and format-dependent failure modes (drift vs inertia) - a ceiling result relevant to RSI loop design.
song2025progco	benchmark	GPT-3.5, GPT-4o, Llama3.1-8B-Instruct on GSM8K, MATH, and IFEval (strict prompt/instruction); self-correction via self-generated verification pseudo-programs (ProgVe) + dual response/program refinement (ProgRe), up to 3 rounds	Accuracy on GSM8K and MATH; IFEval strict prompt and instruction metrics	ProgCo improved GPT-3.5 GSM8K accuracy by +7.28 (to 83.78) and MATH by +5.84 (to 82.34) with 3 self-correction rounds and Python tools, while baseline self-correction methods often degraded scores (e.g., Self-Refine -6.20 and vanilla-reflex -2.20 on GPT-3.5 MATH).	Gains are modest for some benchmarks/tasks in one-round settings; depends on the model's program-generation ability; evaluated on only 3 benchmarks; GPT-4o/Llama gains smaller than GPT-3.5.	moderate	full-text	Evidence that intrinsic self-correction helps only when verification is strong (program-driven): naive self-correction frequently hurts, a key caveat for RSI loops relying on self-feedback.
zhang2024learning	benchmark	LLaMA-2-7B and LLaMA-2-13B fine-tuned on GSM8K CoT plus a checking-correction dataset (~4,500 erroneous-path and ~2,000 correct-path checks generated by LLaMA-7B/13B with GPT-4-1106/GPT-3.5 as feedback models); evaluated on GSM8K, MultiArith, ASDiv, SVAMP	direct-reasoning accuracy and accuracy after self-checking + self-correction	Fine-tuning with the 'Step CoT Check' format raised LLaMA-2-13B GSM8K self-correction accuracy from 45.6% (CoT fine-tuning) to 51.2% (avg +5.5% across datasets; +3.8% for 7B), while base LLaMA models checked reasoning paths at near-chance (~50%) accuracy vs 70% for GPT-3.5 and 89% for GPT-4.	Checking-correction data is generated by stronger models (GPT-4/GPT-3.5) so the loop is not fully self-contained; replacing CoT data with checking data slightly lowers direct reasoning accuracy (worse for 7B); limited to arithmetic-style math word problems.	high	full-text	Shows intrinsic self-correction fails unless the model is explicitly trained to check — evidence that self-improvement loops need learned verifiers rather than naive prompting, a central constraint for RSI.
ganguli2023capacity	benchmark	Anthropic RLHF dialogue models from 810M to 175B parameters; BBQ (58,492 questions, 9 social dimensions), Winogender, and a new racial-discrimination law-school-admission benchmark; interventions Q, Q+IF, Q+IF+CoT; RLHF steps 50-1000	stereotype bias score, gender-pronoun correlation with BLS statistics, racial discrimination rate	The capacity for moral self-correction emerges at 22B parameters and improves with model size and RLHF training: at 175B, Q+IF+CoT reduced overall BBQ bias score by 84% relative to Q-only, and instruction-following could steer pronoun correlations toward 0 or 1 and achieve demographic parity at ~200-600 RLHF steps.	Single model family (Anthropic RLHF); generic prompt interventions; over-correction observed (175B discriminated in favor of Black students by 7% under Q+IF+CoT); authors caution against over-optimism for reducing harm.	high	full-text	Evidence that self-correction capability scales with model size and RLHF — relevant both to safety of self-improving systems and to the claim that bigger models can steer their own behavior.
yuan2024self	benchmark	Llama 2 70B; Open Assistant seed data; 3 iterations of self-rewarding iterative DPO (LLM-as-a-Judge self-rewards); AlpacaEval 2.0, MT-Bench, 9 NLP benchmarks	AlpacaEval 2.0 win rate over GPT-4 Turbo; MT-Bench score; reward-modeling accuracy (pairwise agreement with human judgments)	Self-rewarding iterative DPO improved AlpacaEval 2.0 win rate over GPT-4 Turbo from 9.94% (iter 1) to 15.38% (iter 2) to 20.44% (iter 3), outperforming Claude 2, Gemini Pro, and GPT-4 0613; MT-Bench rose from 6.78 to 7.25 across iterations; adding evaluation fine-tuning raised pairwise reward agreement with humans from 65.1% to 78.7%, and reward-modeling ability kept improving with self-training.	seed data from Open Assistant limits math/code gains (GSM8K drifted 50.72 to 57.70 by iter 3); generations lengthened across iterations (1092 to 2552 tokens), which may inflate win rates; single model family (Llama 2 70B)	high	full-text	Canonical demonstration that an LLM can serve as its own reward model and improve both instruction-following and self-judging ability across iterations - the archetypal self-rewarding RSI loop.
wu2024meta	benchmark	Llama-3-8B-Instruct seed (SFT on EFT); 20,000 prompts (5,000 per iteration x 4 iterations); AlpacaEval 2, Arena-Hard, MT-Bench; judge quality vs Open Assistant human labels and GPT-4 proxy	length-controlled win rate on AlpacaEval 2; Arena-Hard score	Meta-Rewarding (the model judges its own judgments via LLM-as-a-Meta-Judge) raised Llama-3-8B-Instruct's LC win rate from 22.9% to 39.4% on AlpacaEval 2 over 4 iterations and improved Arena-Hard by +8.5% (20.6 to 29.1), beating Self-Rewarding with length control (35.5%) and SPPO (38.77%) without additional human data.	Gains are small on some categories (travel, math); relies on EFT seed data for initial judging ability; judge training uses self-generated meta-judgments that could inherit biases; single 8B model family; 4 iterations only.	high	full-text	Shows that improving the judge along with the actor sustains iterative self-improvement where actor-only self-rewarding saturates - a key mechanism result for RSI.
zhou2025self	benchmark	Mistral-7B-v0.3; iterative DPO over 3 iterations with 4,000 self-generated preference pairs per iteration from Alpaca prompts; evaluated on AlpacaEval 2.0 (win rate / length-controlled win rate), MT-Bench, GSM8K	AlpacaEval 2.0 win rate and length-controlled (LC) win rate; internal reward-model consistency	The two internal reward models (LLM-as-a-Judge vs DPO implicit reward) disagreed on ~50% of preference pairs in self-rewarding training, and enforcing self-consistency (SCIR) raised AlpacaEval 2.0 win rate from 10.8% (M0) to 24.9% at iteration 3 (LC 23.85%), versus 10.7% for vanilla self-rewarding with LLM-as-a-Judge.	Single model family (Mistral-7B) and prompt source (Alpaca); win rates judged by an LLM judge; consistency filtering discards data and may under-use borderline preferences; 3 iterations only.	high	full-text	Diagnoses unreliability in self-generated preference data (a core ingredient of self-improvement loops) and shows consistency constraints on internal rewards substantially improve self-rewarding training.
wang2025temporal	benchmark	Llama-3.1-3B/8B/70B, Qwen2.5-7B, Mistral-7B; AlpacaEval 2.0, Arena-Hard-v0.1, MT-Bench, plus GSM8K, ARC, TruthfulQA, HumanEval; OpenAssistant/UltraFeedback seed data	AlpacaEval 2.0 win rate and Arena-Hard score; chosen-rejected score gap over iterations	Temporal Self-Rewarding (anchored rejection + future-guided chosen) raised Llama3.1-8B's AlpacaEval 2.0 win rate to 29.44% vs Self-Rewarding's 19.69% (+9.75) and Qwen2.5-7B's Arena-Hard to 34.4 vs 21.5 (+12.9) in 2 iterations vs SR's 4, whereas standard Self-Rewarding's chosen-rejected score gap shrank ~9x (vanishing DPO gradient).	Diagnoses Self-Rewarding's representational convergence and gradient collapse rather than fully solving it; uses an extra temporary future-model DPO step; 2 vs 4 iterations makes comparisons compute-matched but confounded; preprint.	moderate	full-text	Shows iterative self-rewarding loops can degrade as chosen/rejected responses converge, and temporal decoupling sustains learning - a self-correction loop stability datapoint for RSI.
tan2025breaking	benchmark	Qwen2.5-Math-7B on DAPO-Math-17K (eval: MATH500, AMC23, AMC24, AIME24, AIME25, HMMT24) and a 375k synthetic arithmetic corpus; cross-model on Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct (BigMath), Qwen2.5-7B-Instruct (WebInstruct-verified/MMLU-Pro); GRPO in Open-R1	Avg@8 and Pass@8 accuracy on reasoning benchmarks; reward-noise/coupling/skew metrics	RLER (ensembled rewards, k=2, with adaptive reward interpolation and disagreement-aware rollout selection) recovered about 96.0% of oracle-reward RLVR test accuracy — an average gain of +45.9% over the pretrained model and +6.2% over the best RLIR baseline (3.6% below RLVR) on DAPO-Math-17K with Qwen2.5-Math-7B — by suppressing the self-confirming loop of over-rewarding high-confidence mistakes.	Main results on a single backbone; k=2 ensemble chosen under a fixed rollout budget; math/verifiable-reasoning domains only; gains over RLVR remain negative (3.6% gap).	moderate	full-text	Diagnoses why self-rewarding RL is unstable (confidence-coupled over-reward creates a self-confirming loop) and provides an ensemble-based fix — central to RSI stability.
fu2026why	theoretical	n/a — formal framework for iterative self-rewarding language model alignment (SRLM): self-reward via model log-probability, DPO-style updates, policy condition number kappa_t; instantiated for linear softmax models	alignment failure rate / error bounds as functions of sample size n and iterations T	Proves a single-step failure lower bound scaling as (kappa_0 log\|Pi\|/n)^{1/2} that depends critically on initial model quality, while the full iterative paradigm improves at O~(1/sqrt(n)) with the dependence on initialization decaying exponentially in the number of iterations T, because the update acts as a contraction on the policy condition number.	Stylized reward (log-probability) and DPO objective; theory on simplified model classes (linear softmax); no experiments on real LLMs; assumptions about prompt distributions and model classes.	moderate	full-text	Provides the first theoretical explanation for why self-rewarding self-improvement converges despite poor initialization — supports stability of iterative self-alignment loops.
xu2024just	benchmark	Mistral-7B and Mistral-Instruct-7B; UltraFeedback prompts (~60k); AlpacaEval 2.0 and MT-Bench evaluation	AlpacaEval 2.0 length-controlled win rate (LC); MT-Bench score	An only-prompting self-rewarding online DPO algorithm (no discriminator/reward model) reached AlpacaEval 2.0 LC win rates of 25.9% (Mistral-7B) and 34.5% (Mistral-Instruct-7B) over 3 iterations, surpassing SimPO by ~4% (and ~2% already at iteration 1), with arithmetic control of the chosen-rejected optimality gap (rejected scores 3->5->7) essential (removal caps LC near 23%).	evaluated on 7B models only; MT-Bench differences within noise; no external replication; prefix-based generation can inject meta-text into outputs	moderate	full-text	Demonstrates a discriminator-free self-rewarding loop for smaller models - self-improvement without external reward models or GPT-4 API feedback.
lee2023rlaif	benchmark	PaLM 2 XS policies, PaLM 2 L and XS AI labelers; Reddit TL;DR summarization, Anthropic HH helpful and harmless dialogue	human win rate, harmless rate, AI-labeler alignment with human preferences	RLAIF matched RLHF (win rates vs SFT: 71% vs 73% summarization, 63% vs 64% helpful dialogue; RLAIF vs RLHF ~50%) and scored the highest harmless rate (88% vs RLHF 76% and SFT 64%); same-size RLAIF beat SFT 68%, and d-RLAIF - where the reward-providing LLM is the exact same checkpoint as the policy - beat SFT 66%, a strict self-improvement instance.	Three tasks only; AI labelers show position bias; the same-size labeler is an instruction-tuned variant rather than the identical checkpoint except in the d-RLAIF helpfulness case.	high	full-text	Foundational evidence that models can improve using AI (including their own) feedback instead of human labels - scalable self-improvement via RLAIF/d-RLAIF.
sun2023principle	benchmark	LLaMA-65b base -> Dromedary via Self-Align (topic-guided red-teaming self-instruct, 16 principles, 5 ICL exemplars, principle engraving, verbose cloning); evaluated on TruthfulQA, BIG-bench HHH Eval, Vicuna benchmark (GPT-4 judged)	TruthfulQA MC1 accuracy, HHH Eval MC accuracy, GPT-4 win rates on Vicuna benchmark questions	Dromedary, aligned with fewer than 300 lines of human annotations (vs 50K+ for InstructGPT/Alpaca), achieved TruthfulQA MC1 accuracy of 69 and HHH Eval overall of 0.83-0.85 (vs ChatGPT 0.87 and Alpaca-65B 0.77), surpassing Text-Davinci-003 and Alpaca on the Vicuna benchmark.	Authors document a 'verbose tax': verbose cloning improved generation quality but hurt multiple-choice benchmarks; principles are exploratory; GPT-4-based evaluation and a modified ranking approach introduce caveats.	moderate	full-text	Self-alignment from scratch: the model generates its own training data under human-written principles with minimal supervision — an alignment-oriented precursor of RSI.
singh2024reinforcemen	survey	n/a (review of RLAIF methods, e.g., for GPT-4-class LLMs)	n/a (qualitative comparison of RLAIF vs RLHF)	Claims Reinforcement Learning from AI Feedback (RLAIF) is a step forward over RLHF because AI-generated feedback scales to more data and is more efficient, using a constitutional preference model for ethical and safe responses, though it may not directly improve understanding of human preferences; no numbers reported.	Review-style claims without experiments; abstract is informal and qualitative.	moderate	abstract-only	Automating the feedback signal (AI feedback) removes the human bottleneck from the alignment loop relevant to self-improvement.
yin2025aligning	benchmark	Gemma-2-9B-it (smaller LLMs in appendix); Persona-Hub personas; AlpacaEval 2.0, MT-Bench, Arena-Hard, Open LLM Leaderboard	AlpacaEval 2.0 length-controlled (LC) and vanilla win rate (WR); MT-Bench score; Arena-Hard WR; Open LLM Leaderboard average	Fully self-synthetic Self-Alignment Optimization (SAO) - prompts, responses, and preferences all generated by the model itself - boosted Gemma-2-9B-it's AlpacaEval 2.0 LC WR by 18.1% and WR by 27.9%, MT-Bench 8.41 to 8.66, and Arena-Hard WR 40.8% to 54.3%, while preserving the Open LLM Leaderboard average (74.28 to 74.41); self-synthesized prompts beat UltraFeedback prompts by 16.46% WR under the same pipeline.	relies on the backbone model's strong self-judging ability (may not transfer to weaker models); main results on a single backbone; using the model's own judgments risks circularity; benchmark-specific	moderate	abstract-only	Shows a model can align itself with fully self-generated prompts, responses, and preferences - a closed-loop, human-data-free self-improvement recipe.
chen2024self	benchmark	zephyr-7b-sft-full (Mistral-7B based) with 50k subset of UltraChat200k; HuggingFace Open LLM Leaderboard, MT-Bench, Big-Bench datasets	Average Open LLM Leaderboard score; MT-Bench score; GSM8k and TruthfulQA accuracy	SPIN self-play fine-tuning improved the base model's average Open LLM Leaderboard score from 58.14 to 63.16 (with 10%+ improvements on GSM8k and TruthfulQA) and MT-Bench from 5.94 to 6.78, outperforming DPO trained with additional GPT-4 preference data while using only the model's own generations from existing SFT data.	Single base model (zephyr-7b); theory guarantees convergence to the SFT data distribution, not beyond it; no exploration/diversity control; results on one model family and mostly leaderboard aggregates.	high	full-text	Direct positive evidence that self-play can convert a weak LLM into a stronger one without new human data - a central result for LLM self-improvement.
wu2024self	benchmark	Mistral-7B-Instruct-v0.2 and Llama-3-8B-Instruct; 60k UltraFeedback prompts (no responses); PairRM preference model (0.4B params); AlpacaEval 2.0, MT-Bench, Arena-Hard, Open LLM Leaderboard	length-controlled win rate vs GPT-4-Turbo on AlpacaEval 2.0	Self-play preference optimization (iterative updates provably converging to the Nash equilibrium of a constant-sum game) achieved a 28.53% LC win rate from Mistral-7B-Instruct-v0.2 and 38.77% from Llama-3-8B-Instruct on AlpacaEval 2.0, outperforming iterative DPO and IPO on MT-Bench, Arena-Hard and the Open LLM Leaderboard without any GPT-4 or stronger-model supervision.	Requires a preference oracle (PairRM) queried each iteration; starts from already well-aligned base models; win rates against GPT-4-Turbo are relative, not absolute quality; scale limited to 7B-8B models.	high	full-text	Game-theoretic self-play alignment with provable Nash convergence, showing iterative self-play improves LLMs without external labels - theory plus strong empirics for RSI.
dong2024self	benchmark	Qwen2-72B/7B, LLaMA3-70B/8B (also Mistral-7B, Mixtral-8x7B, GPT-4); IFEval and FollowBench instruction-following benchmarks; AutoIF self-generated verifiable instructions with execution feedback	instruction-following accuracy (IFEval Loose Acc; FollowBench SSR)	AutoIF self-alignment/strong-to-weak distillation improved SFT, Offline DPO and Online DPO, reaching IFEval Loose accuracy of 88.0% (Qwen2-72B) and 90.4% (LLaMA3-70B) — the first method to surpass 90% — with >5% SSR gains on FollowBench and >4% average gains for the 7B/8B models.	Restricted to instructions verifiable by code; depends on quality of LLM-generated verification functions/unit tests; self-alignment may be bounded by model's own verification ability; results on two benchmark families only.	high	full-text	Shows LLMs can synthesize their own verifiable training data and self-improve without human annotation — a concrete closed-loop self-improvement mechanism with execution feedback.
xiao2025revisiting	benchmark	Llama-3.1-Tulu-3-8B-SFT and Mistral-7B-Instruct-v0.2 (plus Llama-3.1-8B-Instruct) with Skywork-Reward-Llama-3.1-8B; UltraFeedback prompts (~61K); AlpacaEval 2 and Arena-Hard evaluation	AlpacaEval 2 length-controlled (LC) and vanilla win rate (WR); Arena-Hard	The hardest-quartile prompts (bottom 25% by mean sampled reward) underperform equal-size easier subsets in DPO self-play (Tulu LC 24.23% vs 29.53%, WR 18.78% vs 32.09%) and slightly hurt final performance when included, while training on only the ~30% easiest prompts improves overall AlpacaEval 2/Arena-Hard performance; the gap closes at higher model capacity (Llama-3.1-8B hard LC 34.31% vs easier 34.59%).	prompt difficulty is a mean-reward proxy; results on 7-8B models only; curriculum learning and chosen-response-improvement mitigations failed; Arena-Hard gains reported but details less developed	high	full-text	Shows self-play preference optimization gains are prompt-difficulty-dependent - naive scaling of self-play data can yield diminishing or negative returns.
wang2025space	benchmark	Mistral-7B and Zephyr-7B; HuggingFace Open LLM Leaderboard tasks (ARC, GSM8K, HellaSwag, MMLU, TruthfulQA, Winogrande, IFEval, BBH, GPQA, MMLU-Pro); 50k real responses vs SFT's 200k	task accuracy across self-play iterations	Space (NCE-based self-play fine-tuning) improved Mistral-7B GSM8K from 37.68% to 46.02% and IFEval from 23.63% to 35.90% (up to ~10-point gains), and with only 50k real responses outperformed SFT on 200k samples, while the gap-based SPIN baseline degraded after peaking at iteration 2.	SPIN's instability is the motivating diagnosis (objective degenerates when synthetic responses resemble real ones); gap-based variants (S-IPO, S-SimPO) inherit instability; results limited to 7B-scale models and leaderboard-style tasks.	high	full-text	Diagnoses why self-play fine-tuning with gap objectives is unstable and proposes a provably convergent alternative, evidence on limits and stabilization of self-play RSI.
tang2025rspo	benchmark	Over 120 fine-tuned Mistral-7B-Instruct models; self-play preference optimization (SPPO-style) with forward/reverse KL regularization; evaluated on AlpacaEval-2, Arena-Hard-v0.1, MT-Bench, ArmoRM, self-BLEU	Length-controlled win rate (LCWR) on AlpacaEval-2, raw win rate, response diversity	RSPO regularized with a linear combination of forward and reverse KL divergence raised the AlpacaEval-2 length-controlled win rate from 28.5% (unregularized self-play, SPPO) to 35.4% (+6.9% LCWR), with consistent gains on Arena-Hard, MT-Bench, ArmoRM scores and response diversity, while forward KL alone reduced response length and reverse KL alone improved raw win rates.	Single model family (Mistral-7B-Instruct); win rates depend on the GPT-4-based judge; regularization temperature must be tuned; unregularized self-play can over-optimize when the preference model is misspecified.	moderate	full-text	Shows regularization to a reference policy is crucial in self-play alignment — iterative self-training without it over-optimizes; provides convergence guarantees for the regularized game.
jana2026gasp	benchmark	Qwen2.5-Coder-7B trained with guided asymmetric self-play; LiveCodeBench v5 evaluation split (216 questions, 2024.10-2025.02), 601-question training split, 146 goalpost questions; 3 RL seeds	pass@1 and pass@20 on LCB v5	GASP improves LCB v5 pass@20 from 31.15 (unguided asymmetric self-play, AZR) to 33.69 (+2.5%) and pass@1 from 17.49 to 18.26, matching real-data RLVR (33.10), and solves 11 of 146 goalpost questions that no baseline (RLVR or AZR) solves.	Single base model and coding-only domain; pass@1 gains are modest; goalpost solving is intermittent (solutions lie in the tail of the sampling distribution); relies on 3 seeds	moderate	full-text	Goalpost-grounded asymmetric self-play provides a curriculum mechanism for RSI where a teacher generates problems at the student's frontier, with verified gains over unguided self-play.
wilf2025propose	benchmark	PSV-Verus trained by self-play (proposer generating formal Verus/Rust specifications + solver trained by rejection fine-tuning on formally verified solutions); seed corpus from AlphaVerus; evaluated on Dafny2Verus, MBPP, HumanEval; 5 random seeds	pass@1/5/10 on verified code generation	Self-play with formal verification improved pass@1 by up to 9.6x (MBPP 36.78 vs RFT's 3.83; Dafny2Verus 65.63 vs AlphaVerus' 24.06, a 2.73x gain; HumanEval 19.07 vs AlphaVerus' 7.24), with performance scaling with generated-question budget and training iterations and ablations showing formal verification and difficulty-aware proposal are essential.	Verus verifier is sound but not complete, so RFT (not advantage-weighted RL) is used to avoid punishing correct-but-unverified solutions; solving is the most compute-intensive step; limited to the formally verifiable Rust subset.	high	full-text	Positive RSI datapoint: self-play without any human data works in code generation when the correctness signal is sound, contrasting with brittle unit-test-based self-play.
dong2025stp	benchmark	LLM prover + conjecturer trained in self-play (expert iteration on prover, iterative conjecturing on barely-provable statements); Lean and Isabelle; LeanWorkbook, miniF2F-test, ProofNet-test, PutnamBench; 51.3B tokens generated in Lean training	theorem-proving pass rates (pass@3200)	STP proves 28.5% of LeanWorkbook statements, doubling the previous best of 13.2% from expert iteration, and achieves state-of-the-art whole-proof generation on miniF2F-test (65.0%), ProofNet-test (23.9%) and PutnamBench (8/644) at pass@3200.	Very compute-intensive (51.3B tokens); sparse reward plateau partially alleviated but not eliminated; conjecture quality depends on prover capability; results reported at large pass@k budgets; authors note ~98.5% of compute in prior expert iteration yields no successful proofs.	high	full-text	Direct evidence for self-play-based recursive improvement: conjecturer-prover loop keeps improving without new human data — a key RSI mechanism demonstration.
liu2025chasing	benchmark	5 models from Llama and Qwen families; 14 safety benchmarks (HarmBench, WildGuardTest, WildJailbreak, DAN, XSTest, OR-Bench-Toxic, X-Teaming, PAIR, AutoDAN) plus capability evals (ARC-C, GPQA, MMLU, TruthfulQA-MC1, AlpacaEval-2, IFEval)	safety scores across benchmarks; attack diversity (SBERT similarity)	Self-RedTeam, a fully online self-play MARL algorithm co-evolving attacker and defender from a single policy, improved the safety of RLHF-trained models by up to 95% across 14 benchmarks while uncovering more diverse attacks (+17.80% SBERT), generalizing across five Llama/Qwen models, with a Nash-equilibrium-based theoretical safety guarantee.	Numbers read from abstract only; safety gains may trade off against helpfulness (overrefusal seen in baselines); a single reward-model adjudicator could be gamed.	moderate	abstract-only	Self-play attacker-defender co-evolution as an RSI loop that improves safety without external red teams - evidence for adversarial self-improvement.
fu2023improving	benchmark	cohere-command, ai21-jurassic-2, gpt-3.5-turbo, gpt-4, claude-instant-v1.0, claude-v1.3 in a buyer-seller bargaining game with a third LLM critic; ICL-AIF (in-context learning from AI feedback) over multiple rounds	deal price (negotiation outcome) over rounds	Only gpt-3.5-turbo, gpt-4 and claude-v1.3 could continuously improve from iterative AI feedback (e.g., a seller's price improved from $16 to $17 in one example round), while weaker models either failed to understand the game rules or could not incorporate feedback; improvement depended on role, and stronger agents risked breaking the deal when pushing for better prices.	Exploratory, small-scale; in-context learning only (no finetuning); single game domain; deal-breakdown risk under aggressive improvement; conclusions may not transfer to finetuned RL settings.	moderate	full-text	Early evidence of autonomous mutual improvement of LLMs via self-play plus AI feedback (AlphaGo-Zero-style loop), and that a capability/alignment threshold gates who can self-improve.
chojecki2025self	theoretical	n/a - mathematical framework (Generator-Verifier-Updater operator on a Fisher-information statistical manifold) unifying STaR, SPIN, Reflexion, GANs, AlphaZero, GRPO	Expected capability gain kappa; Variance Inequality spectral condition on generation/verification noise	Derives a Variance Inequality: sustained self-improvement (kappa > 0) requires the combined noise of generation and verification to be sufficiently small relative to alignment, curvature, and step size, and in 'diagonal' regimes where the verifier approximates the generator, self-correction typically fails (the Hallucination Barrier) - no empirical numbers are reported.	Purely theoretical with no experiments; relies on regularity assumptions; the proposed empirical kappa-estimation protocol is not run; not peer-reviewed.	low	full-text	Formalizes the verifier-noise condition for stable self-improvement - theoretical backing for the practitioner maxim 'strengthen the verifier, not the generator'.
shumailov2024ai	simulation	LLMs, variational autoencoders (VAEs), and Gaussian mixture models (GMMs) trained on recursively generated data	preservation of the original content distribution (model collapse)	Finds that indiscriminate training on model-generated content causes irreversible 'model collapse' in which the tails of the original content distribution disappear, occurring across LLMs, VAEs, and GMMs, and builds theoretical intuition that the phenomenon is ubiquitous among learned generative models.	Abstract gives no quantitative collapse rates; findings apply to indiscriminate use, with mitigation depending on retaining genuine human data.	high	abstract-only	Key caution for recursive training loops: without fresh human data, self-consumption degrades models irreversibly.
alemohammad2023self	simulation	StyleGAN2 on FFHQ and DDPM on MNIST trained for ~10-20 generations; three autophagous loop families (fully synthetic, fixed real data, biased sampling lambda<1)	FID, precision (quality), recall (diversity) across generations	In fully synthetic loops without enough fresh real data, quality (precision) and diversity (recall) progressively decrease - eventually dropping to zero - while FID steadily increases, a condition the authors term Model Autophagy Disorder (MAD); biased sampling can raise quality but diversity then decreases rapidly.	Image-domain generative models only (not LLMs/text); degradation driven partly by finite sample size; trends saturate rather than fully diverge; theoretical analysis relies on asymptotic approximations.	high	full-text	Foundational negative result on model collapse in self-consuming training loops - the key caution for RSI via synthetic-data bootstrapping.
gerstgrasser2024is	simulation	9M GPT-2 and 12M/42M/125M Llama2 language models on TinyStories; GeoDiff diffusion model on molecular conformations (Drugs data); VAEs on CelebA images; 8 model-fitting iterations	Test cross-entropy / test loss across recursive model-data feedback iterations	Replacing real data with each generation's synthetic data causes model collapse (test error increases with iterations) across all models and modalities, whereas accumulating successive synthetic generations alongside the original real data avoids collapse, yielding equal-or-lower test cross-entropy over iterations.	Small models and datasets (up to 125M params); the data-accumulation setting is 'maximally pessimistic' but still idealized relative to web-scale data dynamics; lower sampling temperature accelerates degradation under replacement	high	full-text	Key evidence that recursive training on self-generated data need not degrade performance if data accumulates, tempering model-collapse pessimism in the RSI debate.
fu2024towards	theoretical	Diffusion models with one-hidden-layer (random-feature) score networks and kernel density estimators trained in self-consuming loops; general, full-synthetic, and balanced data cycles	total variation (TV) distance between learned synthetic distributions and the original real-data distribution across generations	TV distance is controllable when mixed dataset sizes or real-data proportions are large; the full-synthetic cycle requires quartic sample growth or Omega((i-1)/i) real-data proportions to bound error; and expanding synthetic data with fixed real data induces a phase transition where TV distance first rises then declines beyond a threshold.	Simplified models (one-hidden-layer score networks, KDE) and idealized training dynamics; theoretical bounds rather than large-scale empirical validation; specific distributional assumptions.	moderate	full-text	Formalizes model collapse in self-consuming loops and gives conditions (real-data mixing, sample scaling) under which recursive training on own outputs stays stable — directly relevant to RSI feasibility and collapse risk.
ferbach2024self	theoretical	Iteratively retrained generative models on synthetic data and CIFAR10	expected reward of retraining loop and stability of the loop	Proves that when data is curated according to a reward model, iterative retraining acts as an implicit preference optimization mechanism that maximizes expected reward and remains stable when a positive fraction of real data is used at each step, while illustrative experiments on synthetic data and CIFAR10 show the procedure amplifies the reward model's biases.	Results are theoretical and illustrative; assumes curation follows a known reward model and requires some real data each step.	high	abstract-only	Formal result that human/curation feedback in the retraining loop can prevent collapse and steer self-improvement, but amplifies reward biases.
gillman2024self	theoretical	Theory of iterative fine-tuning with correction functions; empirical validation on text-conditioned human motion synthesis (diffusion model) over 50 self-consuming generations	Stability and collapse-avoidance of self-consuming training loops (FID and generation metrics in appendix)	Introducing an idealized correction function makes self-consuming loops exponentially more stable, and a physics-simulator-based self-correction function avoids model collapse on human motion synthesis even at a 100% synthetic-to-real data ratio, where the uncorrected loop collapses within 50 generations.	Theory assumes an idealized correction function; empirical correction relies on expert knowledge (physics simulator) not available in general domains; FID details deferred to appendix	moderate	full-text	Shows a concrete mechanism (self-correction of synthetic data) that stabilizes recursive generative training, relevant to preventing collapse in closed-loop RSI.
kazdan2024collapse	simulation	Multivariate Gaussians, KDE (blobs/circles/moons/swiss roll), Gemma2 SFT on HelpSteer2 (~12.5k examples/generation), linear regression, TinyStories LM pretraining	test loss / negative log-likelihood on held-out real data across generations (and Wasserstein-2 for Gaussians)	The replace workflow (training each generation only on new synthetic data) induces model collapse in all settings - e.g., fit covariance collapses toward 0 and NLL diverges - while the accumulate workflow avoids collapse with bounded test loss; a fixed-compute accumulate-subsample workflow degrades slowly and plateaus, and with <~1024 real examples adding ~1024 synthetic examples improves test loss whereas with ample real data any synthetic data degrades it.	Small-scale settings (2B LM SFT, 12.5k examples) may not extrapolate to frontier pretraining; data-quality filtering is ignored (authors note this is pessimistic for collapse).	high	full-text	Central evidence for the model-collapse debate: whether recursive self-generation degrades models depends critically on the data workflow (replace vs accumulate), shaping RSI feasibility forecasts.
zhao2025convergence	theoretical	n/a (theory: population-level retraining dynamics of self-consuming generative models with heterogeneous Plackett-Luce curators, K candidate pools, reference-mixing weight alpha; generalizes Ferbach et al. 2024)	convergence (KL divergence, total-variation contraction, Hilbert projective metric) and stability under bounded reward perturbations	Proves monotonic improvement of expected exponential reward and KL convergence in pure synthetic retraining; with sufficient reference mixing the finite-pool update is a Banach contraction in TV with geometric convergence and the infinite-pool map contracts in the Hilbert projective metric, while without regularization the dynamics converge but are unstable to bounded reward perturbations.	Population-level idealization (no finite-sample or experimental validation); assumes Plackett-Luce random-utility curation; results are asymptotic and may not capture finite-scale LLM training behavior.	high	full-text	Theoretical account of self-consuming retraining (training on one's own outputs): identifies when recursive retraining converges, and shows reference anchoring is what buys stability — directly relevant to model-collapse debates in RSI.
wei2025self	theoretical	Theoretical analysis plus self-consuming retraining loops of DDPMs on synthetic Gaussian data, CIFAR-10 (VGG11, 92.79% acc) and CIFAR-100 (ResNet56, 72.63% acc); attacks flip 20% of preference pairs	expected reward of generated samples under self-consuming retraining with adversarial curation; robustness conditions of the retraining process	Under self-consuming retraining with adversarially curated data, preference-label-flip attacks (20% of pairs) measurably misalign the retrained model away from true user preferences (gradient-based attacks maintain high efficacy across reward-model architectures; the multi-objective Pareto heuristic achieves the lowest average reward), and injecting real data only partially mitigates the misalignment.	Theory assumes each model update converges to the global optimum of the training objective; attacks require known success ratio kappa; effectiveness is context-dependent; no defense fully prevents misalignment; real-data injection only partially mitigates.	moderate	full-text	Extends self-consuming-loop theory to noisy/adversarial curation: retraining on self-generated, user-curated data is vulnerable to poisoning, bounding the robustness of preference-driven RSI.
wyllie2024fairness	simulation	Simulated supervised-model data ecosystems over many generations, starting from initially unbiased datasets	performance, fairness, and minoritized group representation	Shows through simulation that model-induced distribution shifts over generations cause loss of performance, fairness, and minoritized group representation even from initially unbiased data, and demonstrates that 'algorithmic reparation' interventions that curate representative training batches for SGD can improve fairness relative to other shifts.	Simulation-based with no real-world deployment; no quantitative results reported in the abstract.	moderate	abstract-only	Documents how recursive training loops encode prior mistakes and biases, and a possible corrective intervention.
kim2026why	benchmark	Qwen3-1.7B/8B, DeepSeek-Distill-Qwen-7B, Olmo3-7B-Instruct; DAPO-Math-17k training; OOD eval on AIME24, AMC23, MATH500	math reasoning accuracy (score), response length, epistemic-token counts	Self-distillation (SDPO) in math suppresses epistemic verbalization - solution-guided generation cut epistemic tokens from 182.5 to 8.8 and length from 13,054 to 1,873 tokens - enabling rapid in-domain gains with limited task coverage but causing OOD performance drops of up to 40% on AIME24/AMC23 across four model families.	Controlled study uses 100 selected problems; epistemic-token list is a heuristic; the degradation pattern is specific to math (chemistry shows the opposite trend).	high	full-text	Documents a failure mode of self-improvement: distilling confident, privileged-information teacher traces can quietly destroy generalization - critical caution for RSI loops that train on their own confident outputs.
wang2026denser	benchmark	LLM post-training on Math, Science, ToolUse, Coding domains; SDPO (self-distillation policy optimization) vs GRPO; AIME, Math500, GPQA, ZLogic, MMLU-R, BFCLv4, LCBv6, with teacher-EMA sweeps alpha in {0, 0.5%, 1%, 2%, 5%}	Acc@8 in-domain and OOD benchmarks; retention/forgetting in continual post-training	On-policy self-distillation specialized strongly (Math-SDPO raised AIME from 32.71% to 56.42% and Math500 from 73.85% to 84.29%) but showed stronger forgetting and could even collapse in continual post-training while GRPO preserved prior capabilities, and a fast teacher EMA (alpha=5%) crashed AIME to 34.38% (repaired to 55.00% by restart-and-freeze).	No single teacher-EMA rate is globally optimal; CoT distillation helps ToolUse but hurts Math/Science; denser self-distillation amplifies high-frequency formatting artifacts through a self-reinforcing teacher-student loop; authors conclude on-policy data alone is insufficient for continual learning.	high	full-text	Direct negative evidence that a self-distillation loop (model teaching itself token-by-token) is a fragile, collapse-prone RSI signal in continual settings.
burns2023weak	benchmark	GPT-4-family models spanning 7 orders of magnitude of pretraining compute; 22 NLP classification datasets, chess puzzles (lichess), ChatGPT reward-modeling dataset	Test accuracy; performance gap recovered (PGR) by weak-supervised students	Naive finetuning on weak-model labels consistently outperformed the weak supervisor (PGR typically >20%, often >50% on NLP), and an auxiliary confidence loss recovered nearly 80% of the GPT-2-to-GPT-4 performance gap on NLP tasks, but the ChatGPT reward-modeling setting recovered only ~10% of the gap by default.	No method works consistently across all settings (reward modeling far from full recovery); disanalogies to real superalignment (imitation saliency, pretraining leakage); authors frame results as proofs-of-concept, not deployable solutions.	high	full-text	Quantifies the ceiling of weak-to-strong (self-)supervision: RSI loops inherit supervisor quality limits, and naive self-supervision recovers only part of the capability gap.
lang2024theoretical	theoretical	n/a (theory; illustrative experiments check expansion conditions on finite data)	error bounds for weakly-supervised student models (pseudolabel correction, coverage expansion)	Existing weak-supervision theory cannot account for pseudolabel correction or coverage expansion; the authors derive new expansion-based error bounds showing weak-to-strong generalization occurs when the strong student cannot fit the weak teacher's mistakes without incurring extra error, with expansion checkable from finite data (no performance numbers reported).	Bounds require expansion and robustness-on-average assumptions; empirical validation is limited to checking assumptions rather than end-to-end gains.	moderate	abstract-only	Theoretical foundation for why strong models improve on weak teachers' labels - underpins weak-to-strong supervision as an RSI-adjacent mechanism.
moniri2025mechanisms	theoretical	High-dimensional ridge and weighted-ridge regression (Gaussian features, proportional asymptotics) plus a nonlinear multi-index model with two-layer networks trained by one step of gradient descent; numerical simulations matching the asymptotic predictions	Test-error difference between student and teacher (L_s - L_t) in the high-dimensional proportional limit	Proves three mechanisms of weak-to-strong generalization: (1) a properly regularized student compensates for an under-regularized teacher (lambda_t < sigma^2*gamma_t) and achieves lower test error, while if the teacher is over-regularized the student can never surpass it; (2) a student with regularization structure better aligned to the target (weighted ridge) outperforms an optimally tuned teacher; (3) in a nonlinear multi-index setting a pre-trained student learns easy task-specific features from the teacher while retaining its own hard-to-learn shared features.	Results are asymptotic and restricted to stylized linearized and one-step-GD models; applicability to large-scale LLM fine-tuning is extrapolated rather than demonstrated.	moderate	full-text	Supplies theoretical grounding for weak-to-strong self-training: explains when a stronger student can exceed a weaker teacher's labels and the regularization conditions required, with implications for RSI via synthetic data.
xue2025representati	theoretical	8 datasets: molecular prediction with transformers and 5 NLP tasks; 150 small transformers and 52 LLMs	weak-to-strong generalization (W2SG) performance, formalized as PredGap between weakly- and ground-truth-supervised strong models; label-agnostic representation-overlap metric	Theory plus validation: W2SG is governed by the overlap between the complement of the weak model's principal kernel and the strong model's principal kernel; a label-free metric of this overlap predicts W2SG performance trends across 8 datasets (150 transformers, 52 LLMs), and weak-supervision errors outside the strong model's principal kernel are not replicated even under benign overfitting.	assumes linear heads on fixed representations (NTK-style finetuning); metric correlations rather than causal interventions; molecular/NLP task scope	high	abstract-only	Provides a measurable, label-free predictor of when weak supervision of a stronger model generalizes - directly relevant to superalignment-style RSI.
somerstep2025limitations	theoretical	n/a (theory under a latent-concept-shift transfer framework); toy empirical setting with GPT-4o-mini (strong) and Falcon 7B / Gemma 2B / Llama2 7B / Mistral 8B (weak) on Dolly, TinyTruthfulQA, TinyAlpacaEval	Consistency of weak-to-strong estimators; irreducible error; accuracy and persona-transfer (style) scores	Under latent concept shift, both weak training and label refinement yield inconsistent estimators with irreducible error, whereas a deconvolution-based latent-concept-identification procedure is consistent; empirically, weak training reduced accuracy on TinyTruthfulQA and label refinement left persona bias, matching the theory.	Strong assumptions (algebraic independence, anchor tokens, latent concept shift); the consistent procedure is explicitly impractical for real LLM tasks; toy empirical validation only.	moderate	full-text	Theoretical ceiling on weak-to-strong / self-training approaches: self-generated labels cannot be refined to oracle quality, bounding what RSI via self-labeling can achieve.
yang2024super	benchmark	GPT-2 series, OPT series, Mistral-7B, LLaMA-3-8B/70B, LLaMA-3.1-8B; multi-objective alignment with conflicting targets (helpfulness vs harmlessness); reward modeling and DPO/SimPO preference optimization scenarios	Deception Score (share of 'conflict tax' misalignment occurring in strong-known/weak-unknown regions); alignment accuracy on harmlessness dimension	Weak-to-strong deception exists across all tested settings, intensifies as the weak-strong capability gap increases, and intermediate-model bootstrapping mitigates it only partially (no headline percentage in the read text; deception score is defined as the % of conflict tax in the strong-known/weak-unknown region).	specific multi-objective (helpfulness vs harmlessness) proxy setting; preliminary exploration; bootstrapping mitigation limited; deception scores depend on confidence thresholds	moderate	abstract-only	Cautions that weak-to-strong generalization (basis of superalignment-style RSI) can conceal deception by the stronger model exactly where the weak supervisor is blind.
shin2024weak	benchmark	Qwen1.5 0.5B (weak) supervising Llama3 8B (strong) on 19 datasets (Burns et al./EleutherAI replication); Snorkel label model + 4-layer MLP on 9 WRENCH datasets; synthetic Gaussian mixtures	Weak-to-strong generalization accuracy as a function of overlap density; regret of UCB data selection	Strong-model accuracy improves monotonically as the proportion of 'overlap' points (containing both easy and hard patterns) in the pseudolabeled data increases, approaching the strong ceiling in a high-overlap regime, and the UCB-based overlap-maximizing data-source selection outperforms random sampling (e.g., on Amazon Polarity and DREAM); no single headline accuracy number is reported in the text (results are figure-based).	Simple pattern model only; overlap detection is noisy in low-overlap regimes and can fail; headline results are qualitative trends in figures rather than aggregate numbers; theory requires expansion/robustness assumptions from Lang et al. (2024).	moderate	full-text	Data-centric account of when weak supervision (self-training/pseudo-labeling) can lift a strong model beyond its teacher — conditions under which self-improvement loops can escape the weak-teacher ceiling.
geng2026weak	theoretical	Linear logistic regression; simulations with 80%-accurate student linear models finetuned on pseudo-labels from 70%-accurate weaker teachers (100 teacher draws), plus shared-pretraining simulations (student: 25 epochs on 5000 labeled points; teachers: 10 epochs on 1000-point subset)	0/1 test accuracy and cosine similarity to the true model after SFT on teacher pseudo-labels	In linear logistic regression, SFT on pseudo-labels from a strictly weaker teacher provably improves the student for almost all student-teacher pairs (all 100 randomly drawn 70%-accurate teachers improved an 80%-accurate student), with no capacity mismatch required, though gains are a small but non-trivial constant and cannot reach vanishing error.	Linear one-layer setting only; theory predicts improvement depends inversely on initial student norm while simulations show norm-independence; authors note gains are limited and depend on teacher being weaker by a constant gap	high	full-text	Theoretical grounding that weak-to-strong generalization (a student surpassing a weaker teacher and itself) is 'almost inevitable', lending formal support to the feasibility of RSI loops.
gao2025survey	survey	n/a — systematic survey of self-evolving agents organized by what/when/how/where to evolve (memory, tools, architecture, rewards, single- vs multi-agent, intra- vs inter-test-time), with evaluation metrics/benchmarks and applications in coding, GUI, finance, medical, education	n/a (taxonomy and qualitative synthesis)	No quantitative findings: provides the first systematic review framing self-evolution along three dimensions (what, when, how to evolve), analyzes evaluation metrics and benchmarks (noting coverage gaps and unfair-comparison challenges), and positions self-evolving agents as the path to Artificial Super Intelligence.	Survey only — no new experiments; rapidly changing field; evaluation practices for self-evolving agents identified as limited; safety and co-evolutionary dynamics remain open challenges.	moderate	abstract-only	Maps the self-evolving-agents landscape (mechanisms, stages, evaluation) that the RSI review can use to situate primary evidence.
fang2025comprehensiv	survey	n/a — survey of self-evolving AI agent techniques organized by a unified feedback-loop framework (System Inputs, Agent System, Environment, Optimisers); covers agent prompts, memory, tools, workflows, multi-agent communication, and domain strategies for biomedicine, programming, finance	n/a (taxonomy and qualitative synthesis)	No quantitative findings: the survey catalogs self-evolving agent techniques across framework components and proposes 'Three Laws of Self-Evolving AI Agents' (Endure, Evolve, Cooperate-style safety principles), flagging evaluation, safety and ethics as open challenges.	Survey only — no new experiments; field is nascent and fast-moving; evaluation practices for self-evolving systems noted as immature (coverage gaps, unfair comparison).	moderate	full-text	Provides the conceptual scaffolding for the RSI review: a component-level map of where and how agents self-evolve, plus explicit safety considerations.
zhou2024symbolic	benchmark	GPT-3.5 and GPT-4 language agents; standard benchmarks (HotpotQA hard, MATH, HumanEval) and complex tasks (5 software-development games, creative writing); Agent Symbolic Learning (ASL) with language gradients	F1/EM (HotpotQA), accuracy (MATH), Pass@1 (HumanEval), 1-4 execution score (software), GPT-4 score (creative writing)	Agent symbolic learning raised GPT-3.5 MATH accuracy from 23.8% (hand-crafted Agents baseline) to 38.8% (GPT-4: 56.0% to 60.7%), improved software-development execution scores from 2.4 to 3.8 out of 4, and creative-writing scores from 4.2 to 6.9 (GPT-3.5), by jointly optimizing prompts, tools and pipeline via LLM-generated language gradients, enabling agents to update themselves after deployment.	Proof-of-concept scale; every update requires many LLM calls (cost); performance depends on the initialization of the agent system (over-engineered initializations are unstable); unsupervised (no ground-truth) mode is demonstrated but not rigorously validated against supervised mode.	moderate	full-text	A systematic framework for self-evolving agents (agents as symbolic networks optimized with language back-propagation/gradient descent) — evidence that agent-level self-modification improves task performance data-centrically.
sheng2024language	case-study	GPT-4 (gpt-4-1106-preview), Windows 11 + Python terminal environment, 3 case studies (file-view/edit tools, retrieval augmentation, internet search/navigation)	Qualitative task completion; ability of agent to generate its own augmentations/tools	No quantitative results reported; qualitative case studies show a GPT-4 agent recursively generating its own tools (file viewer/editor, then retrieval, then web search) to solve increasingly complex computer tasks, which the authors describe as a 'recursively self-improving LLM agent'.	Authors state the experiments are tutorial-style, a 'cherry-picked' subset of error-free outputs, and 'not experimentally rigorous'; also flag security vulnerabilities and long-term risks of code-executing self-improving agents.	low	full-text	Early proof-of-concept that LLM agents can self-improve by writing their own scaffolding/tools; evidence is anecdotal but influential for agent-level RSI.
zelikman2023self	benchmark	GPT-4 and GPT-3.5-turbo; LPN task plus 5 transfer tasks (String Grid Distance, Modified Quadratic Assignment, 3SAT, Maxcut, Parity without noise)	meta-utility (average downstream task utility) of improved improver programs	STOP's self-improved improver raised held-out 3SAT solutions from 21.2% to 75.1% and parity-without-noise from 59.3% to 81.7% relative to the seed improver (with monotonic gains over iterations for GPT-4 on LPN), while GPT-3.5 improved in only 12% of runs and mean performance degraded; GPT-4 attempted sandbox circumvention in 0.42% of 10,000 improvement attempts.	Small set of algorithmic downstream tasks; the language model weights are never altered so it is not full RSI; individual runs are non-monotonic and stochastic; safety numbers are attempt-level rates in a synthetic sandbox setting.	high	full-text	First demonstration that an LLM can write code that calls itself to improve itself (recursive scaffolding improvement), plus early quantification of reward hacking / sandbox-bypass frequency in self-improving code.
ishibashi2024can	benchmark	Mistral-7B-family seed models (e.g., SciPhi-Mistral-7B-32k, OpenHermes-2.5-Mistral-7B) merged with LLM-discovered algorithms; evaluated on GSM8k math reasoning	GSM8k accuracy of the seed model after applying self-discovered merging algorithms	The Self-Developing framework's LLM-discovered model-merging algorithms improve the seed model's GSM8k performance by 6% and exceed human-designed approaches like Task Arithmetic by 4.3%, generalizing with 7.4% gains on out-of-domain models without re-optimization.	Demonstrated only in the model-merging domain on a single benchmark (GSM8k); absolute gains are modest; requires verifiable reward signal for algorithm evaluation	moderate	full-text	Evidence for recursive self-improvement of the improvement algorithm itself (LLM autonomously discovering its own merging/fine-tuning methods).
lu2026arex	benchmark	AREX-Turbo (Qwen3.5-4B dense) and AREX-Base (Qwen3.5-122B-A10B MoE); BrowseComp, WideSearch, DeepSearchQA, Humanity's Last Exam (with tools), GAIA, xbench-DeepSearch-2510	task accuracy/success on deep-research benchmarks versus comparable-scale baselines	AREX (inner research loop plus outer self-improvement loop that audits answers constraint-wise and launches targeted follow-up research, with a learned context-update tool and step-aware RL) substantially outperformed comparable-scale baselines across deep-search, wide-search, reasoning and tool-use benchmarks and remained competitive with models using far more activated parameters (specific scores not extracted).	Specific benchmark numbers not read; results on research-style benchmarks may not transfer to general RSI; trained on verified synthetic tasks.	moderate	full-text	Directly an RSI paper: recursive self-improvement of answers via verification-driven refinement loops, evidence that RSI-style agents beat static search.
wang2026metaskill	benchmark	Gemma-4 31B frozen backbone shared by all five pipeline agents (Analyzer, Retriever, Allocator, Proposer, Evolver); OfficeQA, SealQA, ALFWorld with train/validation/held-out splits	held-out test accuracy (ALFWorld: task success rate)	Two-timescale meta-skill evolution (the improvement procedure itself evolved by the same pipeline that evolves task skills) improved held-out accuracy over the raw backbone by +23.54 (OfficeQA), +16.09 (SealQA) and +1.92 (ALFWorld) points, with the slow meta-loop contributing +6.38/+8.05/+1.92 over single-level evolution.	Only three curated benchmarks, so transfer to open-ended real-world tasks is untested; the five-agent pipeline's roles and wiring remain fixed (one-level bounded recursion); meta-update horizon is fixed at H=2 (OfficeQA sheds 9.1 points at H=8); gains on ALFWorld are near-ceiling small.	moderate	full-text	A concrete bounded-recursion RSI implementation - evolving the improvement operator itself - with positive but modest, benchmark-dependent gains.
nguyen2026recursive	benchmark	One shared local backbone (7B on ALFWorld, 134 tasks x 5 seeds; 30B on GAIA, tau-bench, WebShop) with six faithful baselines (ReAct, Reflexion, GEPA, AWM, ACE, Dynamic Cheatsheet); RSEA evolves a three-layer natural-language state (strategy/skills/playbook) with a strict held-out keep-better gate	Task success rate / benchmark score; McNemar paired tests; ablation of the held-out selection gate	RSEA was the strongest single-pass method on ALFWorld (69.3% vs 64.6% ReAct, McNemar p=0.015; 79.4% with retry, best overall) yet no artifact universally won, and unguarded context evolution was unsafe: Dynamic Cheatsheet was near-best on ALFWorld (70.7%) but collapsed on WebShop (0.14 vs 0.43 for ReAct), while removing the held-out gate caused severe overfitting (100% in-sample, a 33-point drop to test).	Single backbone family; four benchmarks only; evolved artifacts are benchmark-specific and transfer is limited; authors note the framing is selection-first, not artifact-first.	high	full-text	Key evidence that the selection gate - not the sophistication of the evolved artifact - determines whether recursive context self-evolution is monotone-safe and free of catastrophic regression.
jin2026exg	benchmark	Multiple sizes of Qwen3 models; HumanEval, EvalPlus, MuSiQue, HotpotQA; baselines include Reflexion and SE-Agent	pass@1/pass@2 on code generation and multi-hop QA; LLM call count and inference latency	EXG (experience graph for self-evolving agents) achieved pass@1 gains exceeding 150% (relative) and pass@2 approaching 30% in the online setting while cutting LLM calls by up to 45.7% and inference latency by up to 30.5% versus reflection- and memory-based self-evolving baselines.	Gain figures are relative and absolute numbers were not read; evaluation uses a two-attempt protocol and Qwen3 models only; retrieval hyperparameters (K=10, fanouts) may not transfer.	moderate	abstract-only	Evidence that non-parametric experience accumulation (graph-structured memory) lets agents improve during deployment without weight updates - a lightweight RSI mechanism.
zhang2026who	benchmark	Frozen single-call LLM solver; MBPP+, Spider 2.0-Snow, and report-generation task families; ten-item anchored reference set; evolutionary metric loop over typed drawback detectors (Double Ratchet co-evolution with the Ratchet skill loop)	evolved-metric agreement with locked hidden ground truth; task lift retained vs ground-truth/rubric grading	The evolved metric gained 0.21 agreement with hidden ground truth on a locked code-generation set that metric selection never read (paired p=0.014), beating the bare LLM judge it contains, and Double Ratchet retained 88-110% of the lift that ground-truth or hand-written rubric grading buys; removing the anchor guards collapsed the metric into a vacuous always-pass detector while removing the detector lifecycle did not.	Only ten anchor items (sparse-anchor regime); validity numbers measured on the code family (MBPP+); task-score is shown to be unable to validate a self-evolved evaluator (a collapsed metric trains skills just as well); small-scale, single-solver setting.	moderate	full-text	Tackles the 'who grades the grader' problem in self-evolving loops: evaluation metrics can themselves be evolved, but anchor discipline (not lifecycle management) is the load-bearing safety ingredient — a Goodhart caution for any self-improving system.
ying2026skilljack	case-study	SkillX and Anything2Skill experience-to-skill pipelines; 150 AppWorld-format trajectories (65 functionally framed poisoned, 65 directly malicious baselines, 20 clean); DeepSeek-v4-flash as extraction/routing/judge model; 4 policy-risk categories	LLM-judge and pattern-detector detection rates on raw trajectories vs extracted skills; attack success rate (routing-level surrogate); persistence after source-trajectory deletion	Skill extraction laundered poisoned experiences: in SkillX, LLM-judge detection dropped from 98.5% for raw poisoned trajectories to 11.4% for derived skills (88.6% evasion vs 27.0% for direct baselines), implanted skills achieved attack success rates of 56.2% (SkillX) and 89.2% (Anything2Skill), and 80.0% of skill-mediated attacks persisted after deleting the original poisoned records.	single model (DeepSeek-v4-flash), not yet replicated across models; trigger rates are routing-level surrogates, not live-compromise rates; detectors are proxy instruments, not certified security controls	moderate	full-text	Reveals the experience-to-skill pipeline of self-evolving agents as a durable attack surface - self-improvement mechanisms can entrench malicious behaviors beyond source-record deletion.
zhao2026safety	cohort	7 LLM backbones (GPT-4o, Claude-4.5-Sonnet, DeepSeek-V3.2, Qwen3-235B-A22B, Qwen3-32B/14B/8B) with AWM and ReasoningBank agent frameworks; experience from WebArena and SafeAgentBench benign tasks; safety evaluated on BrowserART, Agent-SafetyBench, SafeAgentBench	attack success rate (ASR) before/after offline self-evolution and over online self-evolution steps	Experience gathered solely from benign tasks still degraded safety — e.g., GPT-4o ASR on BrowserART rose from 37.0% to 51.0% after offline self-evolution (Claude-4.5-Sonnet 17.0% to 22.0%) — with degradation compounding and persisting online across all backbones (attributed to execution bias reinforcing acting over refusing), while adding refusal experience mitigated ASR but induced over-refusal on benign inputs.	Safety judged automatically by GPT-4o (reported to correlate with human annotations); two agent frameworks and web/embodied domains only; mechanism (execution bias) established via case annotation and length-matched controls rather than full causal intervention.	high	full-text	Core safety evidence for the RSI review: self-evolving agents degrade on safety even from benign experience, exposing a fundamental safety-utility trade-off in experience-driven self-improvement.
cai2025building	framework	StuLife benchmark simulating a student's college journey (3 phases, 10 sub-scenarios); state-of-the-art LLMs including GPT-5	StuGPA score (0-100); task completion, memory utilization, forward transfer metrics	Even the best model, GPT-5, scored only 17.9/100 on StuLife, revealing fundamental deficiencies in long-term memory retention and self-motivated proactive behavior; context engineering (proactive prompting, memory augmentation) improved performance but agents still failed critically.	New benchmark with no external validation; evaluation is simulation-based and narrative-driven; the framework is proposed, not demonstrated to produce self-evolution; single headline score for current LLMs.	moderate	full-text	Provides a measurement instrument (StuLife/StuGPA) for evaluating self-evolving agents - evidence that current LLMs are far from open-ended self-improvement.
qian2025metaagent	benchmark	MetaAgent (QwQ-32B central agent + tool router; also Gemini-2.5-Flash) on GAIA (103 text-only validation questions), WebWalkerQA (680 queries), and BrowseComp (257 questions, Art+History), vs direct reasoning (Qwen2.5-32B, QwQ-32B, GPT-4o, DeepSeek-R1-671B), RAG, workflow agents (Query Planning, Iterative RAG, ReAct, Search-o1-32B), and end-to-end trained WebThinker-32B-Base/RL	Exact Match (GAIA, BrowseComp) and LLM-equivalence accuracy (WebWalkerQA); ablation of reflection/in-house tool components	Weight-free self-evolving MetaAgent improved GAIA average from 22.3 (vanilla QwQ-32B) to 47.6, WebWalkerQA from 4.3 to 52.1, and BrowseComp from 0.0 to 7.1, outperforming workflow baselines (best Search-o1-32B: 39.8/34.1/1.9) and matching or exceeding the end-to-end RL-trained WebThinker-32B-RL (GAIA 48.5, WebWalkerQA 46.5, BrowseComp 2.7) on WebWalkerQA and BrowseComp.	Main results use a single backbone with no reported error bars; BrowseComp/GAIA use subsets; the in-house knowledge base is pre-initialized by simulating three task rounds, and verified reflection uses ground-truth answers (leakage-controlled but training-like).	moderate	full-text	Strong demonstration that self-evolving agents (context + in-house tools, no parameter updates) can match or beat end-to-end RL-trained agents on knowledge discovery - evidence for context-level RSI.
hebbar2026sia	benchmark	Feedback-Agent loop (harness + LoRA weight updates) applied to Chinese legal charge classification (LawBench, 191 classes), Triton GPU kernel optimization (H100), and single-cell RNA denoising	Top-1 accuracy (LawBench), kernel latency in microseconds, mse_norm (denoising)	Combining harness and weight updates (SIA-W+H) outperforms harness-only scaffold iteration on all three domains, achieving 25.1% over prior SOTA on LawBench (70.1% vs 45.0%), 12.4% faster GPU kernels (1,017 vs 1,161 microseconds), and 20.4% better denoising (0.289 vs 0.240 mse_norm).	Requires a task verifier; evaluated on only three domains and one model family; weight updates use LoRA on a single base model	moderate	full-text	Concrete closed-loop self-improvement system updating both agent scaffold and model weights, a direct RSI implementation with consistent cross-domain gains.
wang2026phantom	simulation	Deterministic Counterfactual Fabrication Lab ('MiniArena' board-game micro-lab); LLM-based proposers (glm-5.1, deepseek-v4-pro, kimi-k2.6 among others); 60 runs per arm; byte-exact oracle; planted guard for a castle-rule failure class that provably never occurs	fabrication rate: fraction of runs where the proposer enables the nonexistent-rule guardrail and cites an oracle-refuted violation	In 15/60 runs (vs 0/60 on featureless input) the proposer invented a failure for a provably never-occurring rule and cited a violation the byte-exact oracle refuted, and the effect required three coinciding conditions (rule-shaped pattern, open-ended rule set, failure-presupposing instruction), becoming absorbing under add-only accept loops.	Toy deterministic micro-lab rather than a full agentic workload; single-shot proposal stage plus a stylized add-only accept loop; generalizability to real harness optimization is asserted, not demonstrated; limited proposer sample sizes.	high	full-text	Shows self-improving harness optimizers can hallucinate failures and add phantom guardrails invisible to suppression-only reward - a novel failure mode for RSI loops.
yang2026continually	framework	Llama 3 8B Base; QuALITY niche corpus; Coursera Exam QA; EntiGraph synthetic continued pretraining up to 455M tokens; s1K dataset (1,000 reasoning questions); test-time search over learning-algorithm configurations	closed-book QA accuracy on QuALITY; accuracy scaling vs synthetic token count; downstream benchmark performance	EntiGraph synthetic continued pretraining improved closed-book QuALITY accuracy from 39.49% (Llama 3 8B Base) to 56.22%, scaling log-linearly up to 455M synthetic tokens (adding RAG to the base model gives 39.49% -> 60.35%); the thesis also bootstraps pretraining capabilities from self-generated synthetic data (s1K) and scales test-time search over learning algorithms.	specialized-corpus knowledge acquisition only; synthetic-data factuality constraints; scaling behavior untested beyond 455M tokens; three chapters validated on narrow domains	moderate	abstract-only	A three-pronged program for continually self-improving AI: synthetic data for knowledge acquisition, self-bootstrapped pretraining, and AI-designed training algorithms.
wang2026knowledge	case-study	Generic disposable agents on Claude Haiku 4.5 and GPT-5.4-mini; ARC-AGI-1, ARC-AGI-2, Polyglot, SWE-bench Pro, Terminal-Bench 2 (50-task pools, 10 generations, 3 seeds)	solve rate and dollar cost vs agent-centric baselines (DGM, HyperAgents, Meta-Harness, GEPA, OpenEvolve)	Knowledge-centric self-improvement (agents held fixed; curated knowledge base evolves via task/cross-task forums + distillation) reached 86.7% solve on ARC-AGI-1 ($76) and 68.0% on Polyglot ($126) vs HyperAgents' 70% ($234) and 52% ($190), and 43.8% on Terminal-Bench 2 vs Meta-Harness' 37.6%, with distilled knowledge transferring to held-out tasks and across LLM families.	Controlled case-study on 50-task subsets; baselines rerun under the authors' protocol and margins depend on the LLM; single-run baseline comparisons vs 3-seed own runs; 10 generations may not capture long-horizon behavior; future-dated preprint.	moderate	full-text	Argues RSI progress can be stored in a curated knowledge base rather than agent weights/prompts, a design-axis contribution with cost-efficiency evidence.
aggarwal2026self	case-study	Production deployment: 35+ service microservices platform (~50k LOC), 2 agent interfaces, 11 recorded working sessions, 36 PR reviews across 6 repositories, rule set in version-controlled instruction file	Recurrence rate of ruled-against error classes; rule-set growth; review-effort shift	Across 11 recorded working sessions, accumulated behavioral rules (grown from 5 to 18 rules plus 15+ language standards and a 15-item checklist) achieved a measured 0% recurrence rate for ruled-against error classes, without any model weight updates.	Small n (11 sessions, single organization), no control group, suppression measured only within the observation window (authors note it is not a permanent property), review-focus shift is qualitative/self-reported, non-peer-reviewed arXiv preprint.	low	full-text	Evidence that bounded self-improvement via persistent external memory (accumulated rules from human review) works in production without weight updates; human review remains the signal source, so the loop is human-in-the-loop.
yin2024g	framework	GPT-3.5 as backbone (GPT-4o spontaneously enlisted in unrestricted runs) on coding, science, math (MGSM), and reasoning tasks; monkey-patching-based self-modification of agent's own runtime code	task accuracy vs hand-designed agents and Meta Agent Search (meta-learning optimized agents); convergence speed and cost	Goedel Agent, a self-referential agent that rewrites its own code via prompting guided only by high-level objectives, matched or beat Meta Agent Search on all tasks and outperformed it by 11% on the MGSM math task, converging faster and at lower cost; unrestricted runs exceeded all baselines partly by spontaneously requesting help from GPT-4o.	single backbone model; unrestricted-mode reliance on GPT-4o raises fairness/attribution questions; safety of open-ended self-modification unaddressed; small task set	moderate	full-text	A concrete LLM implementation of the Goedel-machine idea - the agent analyzes and modifies its own code to achieve recursive self-improvement.
wu2025sgm	framework	CIFAR-100 and ImageNet-100 (deep classifiers), CartPole-v1 and LunarLander-v2 (PPO), Rastrigin20 (CMA-ES); 19-100 paired seeds per comparison	certified accuracy/reward improvement via statistical gate (Hoeffding/empirical-Bernstein/e-value lower confidence bounds); accept/reject decisions over irreversible recursive edits	SGM certified a genuine +5.5 pp CIFAR-100 gain under a 30-seed stress test, rejected all ImageNet-100 proposals (e.g., incumbent 76.65% vs proposal 72.62%, mean diff -4.03 pp) and all 8 CartPole regressions (baseline 493.4 vs proposals 447.7), while accepting one LunarLander warm-start edit with mean gain +513.2 (+/-306.9) and certified lower bound +0.04.	proof-based Goedel guarantees relaxed to statistical certificates; assumes bounded paired differences and paired seeds; proposer design and compute tradeoffs left open; evaluated on 4 benchmarks only	moderate	full-text	First statistical safety layer for recursive self-modification - a risk-control gate that certifies edits before commit, foundational infrastructure for safe RSI loops.
zhang2025darwin	benchmark	Coding agents built on frozen Claude 3.5 Sonnet (o3-mini for Polyglot evaluation), 80 self-modification iterations with an archive, SWE-bench Verified and Polyglot benchmarks	percentage of coding tasks solved (pass@1)	The Darwin Godel Machine improved its own coding agent from 20.0% to 50.0% on SWE-bench Verified and from 14.2% to 30.7% on the full Polyglot benchmark over 80 iterations, outperforming baselines lacking self-improvement or open-ended archive exploration, with discovered agents transferring across models (e.g., 19.0% to 59.5% under Claude 3.7 Sonnet) and benchmarks (SWE-bench-evolved agent scored 28.9% on Polyglot vs 14.2% baseline).	Each run takes ~2 weeks with significant API cost; underlying FMs are frozen (no weight-level self-modification); falls short of closed-source SWE-bench SoTA; self-improvement is confined to coding and validated only on coding benchmarks; a run on Polyglot used a different FM for evaluation.	high	full-text	The strongest concrete demonstration of self-referential RSI in LLM systems: the agent edits its own codebase, improving both task performance and its capacity for further self-modification, with archive-based open-ended exploration and sandboxing.
fang2026intelligence	theoretical	n/a (conjecture based on observed AI R&D patterns)	n/a (share of AI-generated artifacts, growth rates)	Proposes the 'Intelligence Flywheel' hypothesis that AI now generates roughly 30% of novel intelligence artifacts, growth rates are shifting toward exponential while total capability remains below human capacity, AI knowledge production outpaces human absorption by orders of magnitude, and the societal impact may parallel the industrial-agricultural transition, so linear projections drastically underestimate near-term disruption.	Stated as a conjecture; the ~30% figure and other observations are asserted without methodology in the abstract.	moderate	abstract-only	Contemporary conjecture quantifying AI's recursive contribution to its own development and arguing for recalibrated governance.
meng2026rsibench	benchmark	4 frontier researcher agents (Claude Code Opus-4.8/Sonnet-5, Codex gpt-5.6-sol/terra) iteratively revising LoRA SFT training-data strategies for a fixed Qwen/Qwen3.5-35B-A3B-Base target on 6 benchmarks (SWE-bench Verified/Multilingual/Pro, Terminal-Bench 2.0, GPQA Diamond, AIME 2026); fixed Tinker/Harbor-E2B stack, 16h/$500 budgets, one representative run per setting (24 settings)	official benchmark-native score of selected checkpoint; improvement over first valid attempt; within-run selection trajectories	Agents improved on their first valid attempt in 58.33% of settings (14/24) but feedback-driven improvement was rarely monotonic - among 23 searches continuing after their best observed score, 78.26% (18) finished with a lower-scoring final attempt and 5 only recovered the peak; best official scores were still weak on hard tasks (9.00% SWE-bench Pro, 22.00% SWE-bench Multilingual, 20.22% Terminal-Bench 2.0), and an early same-family RSI experiment improved its candidate from 8% to 21% on SWE-bench Pro but stayed below the 33% unadapted-model reference.	One representative run per agent-benchmark setting (no repeated trials), checkpoint selection and official evaluation use the same task subset so adaptive generalization to held-out tasks is not established, and researcher identity bundles the LLM, scaffold, and reasoning effort.	moderate	full-text	Provides benchmark evidence of a 'discovery-reliability gap': current frontier agents can make useful data-centric discoveries for RSI but cannot yet convert feedback into consistent, monotone improvement.
xue2026past	benchmark	PAST-Bench: 26 task-family scenarios, 204 episode tasks, 4 capabilities (Memory, Procedural, Information gathering, Update); 7 base models (GLM-5.1, Kimi K2.6, DeepSeek-V4-Pro, MiniMax-M2.7, GPT-5.4, Claude Sonnet 4.6, etc.); 4-5 agent frameworks (Hermes, Hermes+, nanobot, ZeroClaw, Agent-Zero)	self-evolution gap Delta (persistence-on minus persistence-off task score, 0-1 scale) plus mechanism-evidence score (Mech) for pathway attribution	Retained experience improved later-task scores for every base model (Overall Delta +0.13 to +0.24; e.g., GPT-5.4 +0.24, GLM-5.1 +0.20, Kimi K2.6/DeepSeek-V4-Pro +0.17) but gains were capability-specific and equal headline gaps hid different mechanism evidence; Hermes+ raised Overall Delta from +0.13 to +0.15 with Mech 0.73 vs 0.64 and doubled the Update gain (+0.12 to +0.24) on MiniMax-M2.7, while Agent-Zero regressed (Delta -0.08).	synthetically constructed task families evaluated in isolation; run-to-run variation (Overall 0.13+/-0.04 vs 0.15+/-0.06) exceeds the Hermes+ gain; effects remain capability- and model-dependent	high	full-text	First benchmark isolating whether retained experience actually makes agents improve across sessions, with matched persistence on/off controls - measurement foundation for RSI in personal agents.
sahoo2026sahoo	benchmark	Qwen3-8B improved over up to 20 self-improvement cycles on 189 tasks (63 HumanEval code, 63 TruthfulQA truthfulness, 63 GSM8K math), with SAHOO safeguards (Goal Drift Index multi-signal detector, constraint preservation checks, regression-risk quantification); thresholds calibrated on 18 tasks x 3 cycles; ~2000 bootstrap samples for CIs	Task quality (pass@1 for code, accuracy for truthfulness, exact match for math), Goal Drift Index vs 0.44 threshold, constraint preservation score (CPS), capability-alignment ratio	Safeguarded recursive self-improvement produced quality gains of +18.3% on code (0.672 to 0.795), +16.8% on math (0.689 to 0.805), and +3.8% on truthfulness (0.678 to 0.704), with mean GDI 0.335 (below 0.44), perfect constraint preservation in code and math, 170 truthfulness violations (53.5% fabrication), 173/189 tasks converging (mean 8.8 cycles), and regression events of 4.5% (0.7% excluding one outlier task).	Single base model (Qwen3-8B); GDI/CAR metrics are self-defined and not externally validated; thresholds calibrated on a small 18-task set; the paper discloses LLM-assisted manuscript preparation and code development, and authors note calibration is specific to benchmark distributions and decoder-style models.	moderate	full-text	Makes alignment drift during RSI measurable (drift index, constraints, regression risk) and shows early improvement cycles are cheap while later gains carry rising alignment costs - direct evidence for the capability-alignment frontier.
chen2026recursive	survey	1,250 arXiv papers (2024-2026) on self-refine, self-reward, self-play, self-evolve, and auto-research	n/a (two-axis taxonomy: what is improved x degree of loop closure; verification hierarchy)	The survey finds that demonstrated self-improvement strength tracks a verification hierarchy from formal verifiers (strongest) to intrinsic self-assessment (weakest), that failure modes (self-confirming loops, model collapse, diversity collapse) follow from violations of it, and that open-ended recursive self-improvement remains bounded by grounding requirements, collapse dynamics, and compute constraints on every measured axis.	Scope limited to 2024-2026 arXiv literature; taxonomy and hierarchy assignments are judgment-based; no quantitative meta-analysis of effect sizes.	moderate	full-text	Provides the organizing taxonomy for the RSI review (bounded self-refinement vs open-ended RSI) and the verification-hierarchy lens for interpreting empirical self-improvement results.
zhang2026self	theoretical	n/a (formal argument via von Neumann's complexity threshold and Kleene's Second Recursion Theorem, plus a review of empirical LLM metacognition/self-evaluation literature)	n/a (conceptual: existence of introspective programs; structural bottlenecks to self-simulation)	Argues sustainable RSI requires an 'introspection threshold' analogous to von Neumann's complexity threshold — introspective programs provably exist (Kleene's Second Recursion Theorem), but current LLMs exhibit only quasi-introspection and cannot cross the threshold because of missing complete self-access, the feedforward Transformer architecture, and computational-class limits that prevent fixed-point iteration.	No new experiments; the complexity/introspection threshold is conceptual rather than quantified; relies on a selective reading of empirical self-correction results (saturation, degradation without external feedback, model collapse).	moderate	full-text	Provides the theoretical frame for when autonomous self-improvement is sustainable rather than degenerative, connecting RSI to self-reference, computability theory, and model collapse.
hughes2024open	framework	n/a (position paper; examples analyzed include AlphaGo, AdA, POET, and foundation models)	n/a (formal definition of open-endedness as continuous generation of artifacts that are both novel and learnable with respect to an observer)	No empirical numbers - the paper argues open-endedness is an essential property of artificial superhuman intelligence and that combining open-ended algorithms (RL, self-improvement, task generation, evolution) with foundation models is on the critical path to ASI, while cataloguing safety risks of open-ended systems.	Position paper with no empirical validation; definition is observer-relative and not operationalized into a benchmark	low	full-text	Provides the conceptual framing linking open-endedness and self-improvement to ASI, useful for the review's theoretical framing of RSI.
faldor2024omni	simulation	Foundation models (LLMs/VLMs) generate Python task + environment + reward code; R2D2 robot in simulated worlds; PPO RL agents trained per task; long run (200 iterations, simulated learning) and 5 short runs with real RL training	task diversity (cell coverage), open-ended progress (ANNECS-OMNI), task learnability	In a 200-iteration run OMNI-EPIC generated diverse, increasingly complex tasks far beyond its seeds; in RL runs agents completed 16 tasks and failed 6 (1 uninteresting), and OMNI-EPIC significantly outperformed controls (without archive or without models of interestingness) on both cell coverage and ANNECS-OMNI (p<0.05, Mann-Whitney U), with 72.7% human agreement with its success detector (n=50).	Not yet Darwin Complete — constrained by the chosen simulator; short RL runs due to compute; specialist agents per task with no demonstrated cross-task generalization; success detector (LLM) imperfect.	moderate	full-text	Demonstrates an open-ended self-improvement loop: models autonomously generate curricula (tasks + rewards) matched to agent ability — a step toward AI-Generating Algorithms and endless capability growth.
zhang2023omni	benchmark	PPO agents in Crafter, BabyAI (MiniBossLevel) and AI2-THOR kitchen; foundation-model Models of Interestingness (GPT-3 Davinci, GPT-4); 10 seeds, 100M (Crafter/BabyAI) and 1M (AI2-THOR) timesteps	average task success rate and number of tasks learned (success >= 0.2, or >= 0.6 in AI2-THOR)	OMNI learned 82 tasks in Crafter (14 interesting) vs 55 for learning-progress and 4 for uniform sampling, with average success 0.56 vs 0.42 and 0.030, and in the infinite AI2-THOR task space learned 13 tasks vs 2 (LP) and 0 (uniform), performing near-parity with an oracle model of interestingness.	FM-based interestingness is a proxy for human judgment (Goodhart risks acknowledged); small number of domains; 2D/3D RL settings rather than LLM benchmarks; results are medians over 10 seeds with bootstrap CIs.	moderate	full-text	Contributes auto-curriculum — an AI selecting its own next task to learn via a learned notion of interestingness — as a building block for self-improving AI and AI-Generating Algorithms.
shinn2023reflexion	benchmark	LLM agents (GPT-4-class) on HumanEval, sequential decision-making, and language reasoning tasks	pass@1 accuracy on HumanEval	Reflexion, which improves agents through verbal self-reflection stored in episodic memory rather than weight updates, achieves 91% pass@1 on HumanEval, surpassing the previous state-of-the-art GPT-4 at 80%.	Gains come at inference/trial time rather than as durable weight updates; performance varies with feedback signal type and agent.	high	abstract-only	Weight-free self-improvement via verbal feedback, showing LLM agents can learn from their own mistakes without fine-tuning.
bai2022constitutional	benchmark	Anthropic LMs (810M-52B); 182,831 red-team prompts; 16 constitutional principles; RLHF vs RLAIF comparison; 10,274 helpfulness + 8,135 harmlessness crowdworker comparisons	Crowdworker Elo (helpfulness/harmlessness); preference-model harmfulness scores; absolute harmfulness (0-4)	RL-CAI, trained with AI-generated feedback (RLAIF) from a 16-principle constitution and no human harmlessness labels, was significantly more harmless than human-feedback RLHF models at matched helpfulness, with harmlessness improving monotonically with the number of self-critique/revision steps and RL-CAI virtually never evasive.	Helpfulness labels still human-supplied; PM scores are poorly calibrated at extremes; over-training leads to Goodharting (boilerplate 'you are valued' responses); principles chosen ad hoc; absolute harmfulness scale not well-calibrated.	high	full-text	Early large-scale demonstration of self-improvement via AI feedback replacing human labels - key evidence for scaling supervision in alignment.
christiano2018supervising	framework	Algorithmic tasks (permutation powering, expression evaluation, union find, shortest path, wildcard search); autoregressive MLP + Transformer; hand-coded decomposition oracle H	Accuracy on tasks at increasing recursion depth; sample complexity (oracle calls)	Iterated Amplification learned the algorithmic tasks with at most a modest slowdown vs supervised learning from ground truth (~2x computation per question) and required only tens of thousands of oracle queries (7,000-24,000 per task) versus tens of millions of ground-truth examples for direct supervised learning.	Toy algorithmic tasks with hand-coded decompositions (human replaced by an algorithm); no evidence humans can decompose real-world tasks; instantiated only with supervised learning, not the RL variant; accuracy of the student always trails the amplified expert by design.	moderate	full-text	Foundational scalable-oversight proposal: build training signals for hard tasks by amplifying weak experts on easier subproblems - ancestor of recursive reward modeling in self-improvement supervision.
leike2018scalable	framework	n/a (research direction, conceptual)	n/a (no experiments)	No empirical results; proposes reward modeling plus recursive reward modeling as a research direction in which agent A_{k-1} assists the user in evaluating outcomes when training A_k, letting users train agents in domains they cannot evaluate directly (no numbers).	Explicitly a research agenda, not a solution; depends on two assumptions (learnable intentions; evaluation easier than production) and on bounding error accumulation across the recursion.	moderate	full-text	Seminal framework for recursive reward modeling - a canonical RSI architecture (agents training better agents via amplified evaluation).
li2023selfalign	benchmark	LLaMA 7B/33B/65B (plus LLaMA2-70B); 3.2k OpenAssistant seed + 502k Clueweb segments; AlpacaEval (805 prompts), human eval, MMLU	AlpacaEval win rate vs text-davinci-003 (GPT-4 judged) and human preference	Humpback (instruction backtranslation: self-augmentation + self-curation, 2 iterations) achieved 79.84% (33B), 83.71% (65B) and 87.94% (70B) win rates using only 3k human-annotated plus ~45k self-curated examples, outperforming all non-distilled LLaMA models (Guanaco 65B: 71.80%) and rivaling distilled ones (Vicuna 33B: 88.99%); uncurated augmented data gave no gains while curated data scaled with coefficient alpha=6.95.	Depends on a web corpus and model self-scoring for curation; GPT-4-judged win rates carry known biases; authors flag bias and safety concerns.	high	full-text	Demonstrates an iterative self-training loop (generate -> self-curate -> retrain) that improves instruction following without external distillation - direct RSI evidence.
chen2023selfdebug	benchmark	code-davinci-002, gpt-3.5-turbo, gpt-4, StarCoder on Spider (text-to-SQL), TransCoder (C++-to-Python), MBPP (text-to-Python)	Execution-based accuracy on code generation benchmarks	Self-Debugging with code explanation consistently improved baseline accuracy by 2-3% on Spider (and 9% on the hardest query level), and by up to 12% on TransCoder/MBPP when unit tests were available, while matching or outperforming baselines that sampled more than 10x candidate programs.	Gains are modest without unit tests (2-3%); requires an execution environment; few-shot prompting only (no weight updates); effectiveness varies by model.	high	full-text	Evidence for bounded self-refinement in code generation via execution feedback and self-explanation - the 'code self-repair via execution feedback' category of RSI.
guo2024online	benchmark	PaLM 2 S/L on TL;DR summarization, Helpfulness, and Harmlessness tasks; Gemini Pro and human evaluation	Win rate against SFT baseline and against offline counterparts	Online AI feedback (OAIF) with on-policy DPO/IPO/SLiC outperforms offline versions and RLHF: online DAP methods achieve an average ~66% win rate (range ~64-71%) against their offline counterparts in human evaluation, with online DPO surpassing offline DPO's win rate after 4,000 training steps.	Some win rates judged by an LLM (Gemini Pro) rather than humans; evaluation tasks limited to summarization and helpfulness/harmlessness	moderate	full-text	Evidence that on-policy (online) feedback - sampling from the current model - materially improves self-training alignment, a key design choice for RSI loops.
huang2023cannot	benchmark	GPT-3.5-Turbo, GPT-4, GPT-4-Turbo, Llama-2-70b-chat on GSM8K, CommonSenseQA, HotpotQA (200-question subsamples for larger models)	Accuracy before vs after intrinsic self-correction (no oracle labels or external feedback)	Without oracle labels, intrinsic self-correction consistently decreases accuracy across all models and benchmarks - e.g., GPT-4-Turbo GSM8K drops from 91.5 to 88.0 after round 1 - and GPT-3.5 retains its initial GSM8K answer only 74.7% of the time, flipping correct answers to incorrect more often than the reverse.	Subsampled evaluation (200 questions) for GPT-4/Llama-2; conclusions depend on prompt choices for initial generation and correction	high	full-text	Critical negative result: intrinsic self-correction without external feedback fails, a key boundary condition for RSI claims about self-correction.
gulcehre2023rest	benchmark	Machine translation models on IWSLT 2014 De-En, WMT-2020, and Web Domain En-Zh, scored with MetricX reward model (BLEURT/BLEU also reported)	Average reward (MetricX) on validation set; BLEURT/BLEU	Each Improve step improves translation quality on all three datasets, and a second Grow step adds a further 5.3 points on IWSLT 2014 De-En and 0.8 points on Web Domain En-Zh; the best ReST variant with N<10 best-of-N sampling matches the BC baseline with N=200.	Depends on a learned reward model (MetricX) that is itself vulnerable to reward hacking; evaluation limited to machine translation	high	full-text	Canonical growing-batch self-training loop (sample-filter-finetune, repeat) demonstrating measurable iterative self-improvement in LLMs.
tao2024survey	survey	n/a (literature survey of self-evolution methods for LLMs and LLM-based agents)	n/a	No quantitative finding; the survey organizes the field into a 4-phase iterative cycle (experience acquisition, experience refinement, updating, evaluation), categorizes autonomy into low/medium/high levels (most current frameworks are low-level), and reports that current methods struggle to improve after more than three rounds of self-evolution, with model collapse, stability-plasticity, and superalignment flagged as open problems.	Survey without new experiments; claims about performance ceilings and collapse rely on the cited literature; taxonomy is descriptive rather than evaluative.	high	full-text	Provides the organizing taxonomy and open-problem list (autonomy levels, three-round improvement ceiling, model collapse, superalignment) that structures the RSI review.
shumailov2023curse	simulation	OPT-125m fine-tuned on wikitext2 across generations of model-generated data (5-way beam search, 64-token blocks); GMMs, VAEs, 1D Gaussians; theoretical toy models	Perplexity on wikitext2; Wasserstein-2 distance / variance divergence of recursive estimation	Sequentially fine-tuned OPT-125m degraded from 20 to 28 perplexity on wikitext2 when trained on its own generations with no original data preserved, while preserving 10% original data yielded only minor degradation, and theory shows estimation variance grows linearly (Var = sigma^2(1+n/M)) so risk diverges unless sampling scales superlinearly.	LLM experiments limited to fine-tuning small OPT-125m (authors note only 'early signs' of collapse appear in fine-tuning vs from-scratch training); GMM/VAE results are synthetic; real-world severity could differ in either direction.	high	full-text	Canonical negative result for recursive self-improvement: training on self-generated data causes irreversible model collapse (tail loss), foundational for RSI risk analysis.
voyager2023	benchmark	Voyager (GPT-4, blackbox, no fine-tuning) in Minecraft via Mineflayer with automatic curriculum + growing skill library + iterative prompting; baselines AutoGPT, ReAct, Reflexion; 3 trials; zero-shot generalization in a new world	Unique items discovered, tech-tree milestone speed, distance traversed, zero-shot task success rate	Voyager discovered 63 unique items (3.3x more than baselines) within 160 prompting iterations, unlocked the wooden/stone/iron tech-tree levels 15.3x/8.5x/6.4x faster and was the only method to reach diamond (102 iterations, 1/3 trials), traversed 2.3x longer distances, and solved all 4 unseen zero-shot tasks (3/3 trials) where baselines solved none; removing self-verification cut item discovery by 73% and a random curriculum by 93%.	GPT-4 API cost (~15x GPT-3.5); text-only perception; occasional hallucinations (impossible crafting tasks); 3 trials per condition; blackbox queries rather than weight updates.	moderate	full-text	Demonstrates compounding, compositional skill-library growth in a lifelong-learning embodied agent — positive agent-level evidence for RSI-like accumulation without parameter updates.
olausson2023silver	benchmark	CodeLlama-13b-instruct, GPT-3.5, and GPT-4 on Python tasks from HumanEval and APPS (300-task subset); self-repair trees with bootstrapped pass-rate estimates (N_p=50, N_t=1000) at matched sample budgets vs i.i.d. no-repair baselines	pass@k with repair cost accounted (pass rate at equivalent total sample budget); repair success rate	Self-repair is not a silver bullet: gains were modest and inconsistent - up to ~8% above the i.i.d. baseline for GPT-4 on APPS (up to ~34% relative gain on competition-level problems for GPT-3.5) and up to ~10% for Code Llama on HumanEval, often at or below baseline at small budgets; spending budget on diverse initial samples beat extensive repair (10 initial + 1 repair gave 1.05x pass@20 vs 0.97x for 2 initial + 10 repairs), and boosting feedback quality unlocked the gains (human feedback raised GPT-4 repair success from 33.3% to 52.6%, a 1.58x factor).	Assumes executable unit tests are available; results vary strongly with difficulty subset and budget; feedback-bottleneck hypothesis supported but not fully causal.	high	full-text	Central conditional evidence on self-repair for RSI: the model's own feedback is the bottleneck, so self-correction gains are limited unless feedback quality is externally boosted.
alphaevolve2025	case-study	Evolutionary coding agent using Gemini 2.0 Flash/Pro ensemble; 14 matrix-multiplication problems, 50+ math construction problems, 4 Google engineering problems (Borg scheduling, Gemini kernel tiling, TPU arithmetic circuits, attention)	Quality of discovered algorithms (e.g., number of scalar multiplications), % of problems where SOTA was surpassed/matched, kernel speedup	AlphaEvolve improved SOTA for 14 matrix multiplication algorithms, matched best-known constructions on ~75% and surpassed them on ~20% of 50+ math problems (including 4x4 complex matmul in 48 multiplications, the first improvement over Strassen in 56 years), and discovered a Gemini kernel tiling heuristic giving 23% average speedup and ~1% reduction in Gemini training time.	White paper without full peer review; requires machine-gradeable evaluation metrics so manual-experimentation tasks are out of scope; results are self-reported with no independent replication; compute costs of evolution not fully disclosed.	moderate	full-text	Evidence that LLM-driven evolution grounded in execution can improve critical infrastructure, including the training of the LLM underpinning AlphaEvolve itself - a partial self-referential closed loop.
sica2024	benchmark	SICA self-improving coding agent (Claude Sonnet 3.5 v2 + o3-mini reasoning sub-agent), 50-problem random subset of SWE-Bench Verified, 50 LiveCodeBench, 2 synthetic benchmarks, 15 iterations, ~$7,000 API cost	Accuracy on SWE-Bench Verified and LiveCodeBench; utility combining score/cost/time	Self-referential editing of its own Python codebase raised SWE-Bench Verified accuracy on a 50-problem subset from 17% to 53% over 15 iterations (0.17 -> 0.53), with smaller gains on LiveCodeBench (0.65 -> 0.71) and negligible gains on AIME/GPQA where o3-mini alone scored 87%/79% vs the agent system's 76% average.	Single run on a random 50-problem subset; path-dependent quality ('taste') of self-proposed changes raises variance; expensive (~$7k); self-improvement saturates when the base LLM is already strong; 5-minute timeouts depress initial scores.	moderate	full-text	Direct empirical demonstration of scaffold-level recursive self-improvement in a coding agent, with clear saturation when the underlying model already performs well.
wang2022selfconsistency	benchmark	UL2-20B, LaMDA-137B, GPT-3-175B (code-davinci-001/002), PaLM-540B; GSM8K, SVAMP, AQuA, MultiArith, ASDiv, StrategyQA, ARC-challenge, CSQA	accuracy vs chain-of-thought greedy decoding	Sampling diverse reasoning paths and marginalizing by majority vote improved CoT accuracy on PaLM-540B by +17.9 absolute on GSM8K (56.5 to 74.4), +11.0 on SVAMP, +12.2 on AQuA, +6.4 on StrategyQA and +3.9 on ARC-challenge, with gains growing with model scale.	Applies only to tasks with a fixed answer set; requires sampling many paths (compute cost); model probabilities are poorly calibrated, so unnormalized weighting performs worse than simple majority vote.	high	full-text	Shows a purely inference-time self-ensemble (no weight updates) yields large reasoning gains, a zero-parameter form of self-improvement relevant to RSI scaling.
gou2023critic	benchmark	ChatGPT, text-davinci-003, LLaMA-2 (7B/13B/70B) on free-form question answering, mathematical program synthesis, and toxicity reduction	F1 on QA, accuracy on math tasks, toxicity probability	CRITIC's tool-interactive verify-then-correct loop improved ChatGPT by 7.7 F1 across three QA tasks, 7.0% absolute accuracy on three mathematical reasoning tasks, and reduced toxicity probability by 79.2%, while self-correction without external feedback yielded only modest gains or deteriorated performance.	Requires access to text-to-text tool APIs; authors find all tested LLMs are unreliable at validating their own outputs, so gains depend on external verification	high	full-text	Distinguishes effective (tool-grounded) from ineffective (intrinsic) self-correction, establishing that external feedback is crucial for consistent LLM self-improvement.
haluptzok2024language	benchmark	Three GPT-Neo models (125M-2.7B) fine-tuned on self-synthesized programming puzzles (code-based format) filtered for correctness by a Python interpreter; Codex-generated synthetic code also used	Test accuracy on held-out human-authored programming puzzles	LMs fine-tuned on their own interpreter-verified synthetic puzzles and solutions more than double their test accuracy on held-out human-authored puzzles, demonstrating self-improvement via self-generated verified data without human-authored problems.	Restricted to the programming-puzzle format (expressible but specialized); requires an interpreter for verification; gains measured on puzzles similar in distribution to the synthetic curricula	high	full-text	Direct demonstration of an LM teaching itself to program better through self-play with machine verification - a canonical RSI result.
zheng2023judging	benchmark	GPT-4 as judge; MT-Bench multi-turn questions; Chatbot Arena crowdsourced battles; LLaMA/Vicuna variants	agreement between LLM judge ratings and human preferences	Strong LLM judges such as GPT-4 match both controlled and crowdsourced human preferences with over 80% agreement, the same level as agreement between humans, making LLM-as-a-judge a scalable and explainable proxy for human preferences.	Identifies position, verbosity, and self-enhancement biases plus limited reasoning ability of judges, and proposes only partial mitigations.	high	abstract-only	Enables scalable automated evaluation, a critical feedback component for closed-loop self-improvement pipelines.
gao2023scaling	benchmark	Synthetic setup in an InstructGPT-style environment: fixed 'gold-standard' reward model labels proxy reward models (12M-3B parameters) trained on human preference data; policies 1.2B-6B; optimization via PPO RL or best-of-n sampling	gold reward model score as a function of KL divergence from the initial policy	Gold reward rises then falls with optimization, fitting R(d)=d(alpha-beta*d) for best-of-n and a similar form for RL; coefficients scale smoothly (approximately logarithmically) with proxy RM parameter count, RL is far less KL-efficient than best-of-n, a KL penalty raises proxy reward without measurably improving gold reward, and larger policies benefit less from optimization without overoptimizing more.	Gold reward model substitutes for humans (synthetic labels); single environment; functional forms may not hold near the origin; KL-penalty result flagged as potentially hyperparameter-sensitive.	high	full-text	Quantifies the Goodhart/overoptimization ceiling when optimizing against a learned proxy reward — a fundamental limit for self-improvement loops that optimize self-generated or proxy rewards.
yang2023opro	benchmark	LLMs (PaLM-class) as optimizers on GSM8K, Big-Bench Hard, linear regression, and traveling salesman problems	task accuracy of LLM-optimized prompts	Optimization by PROmpting (OPRO), where the LLM iteratively generates new solutions from prompts containing previously evaluated candidates, produces prompts that outperform human-designed prompts by up to 8% on GSM8K and up to 50% on Big-Bench Hard.	Improvements are prompt-level rather than model-level; gains vary across the LLMs used as optimizers.	high	abstract-only	Demonstrates LLMs optimizing their own instructions, a direct self-optimization loop at the prompt level.
aiscientist2024	case-study	GPT-4o-based pipeline (The AI Scientist); 3 ML subfields (diffusion modeling, transformer language modeling, learning dynamics); 500 ICLR 2022 OpenReview papers for reviewer validation	Automated reviewer agreement with human decisions (accuracy, F1, AUC); generated-paper acceptance scores; cost per paper	The automated reviewer achieved near-human performance (70% accuracy; 65% vs 66% balanced accuracy on balanced data; F1 0.57 vs 0.49 human; AUC 0.65 both) and The AI Scientist produced full research papers at under $15 each that exceeded the acceptance threshold of a top ML conference as judged by its own reviewer.	Papers are judged by the system's own automated reviewer (higher false-positive rate 0.31 vs 0.17 human) with no external validation of paper quality; some models needed score thresholding due to over-optimism bias; reviewer is part of the same loop it evaluates.	moderate	full-text	Canonical demonstration of closed-loop AI research automation ('auto research' category in RSI); shows evaluator quality, not generation, is the bottleneck in automated discovery loops.
wang2023mathshepherd	benchmark	Mistral-7B and other open-source LLMs on GSM8K and MATH	accuracy on GSM8K and MATH	Math-Shepherd, an automatically trained process reward model, improved Mistral-7B via stepwise PPO from 77.9% to 84.1% on GSM8K and 28.6% to 33.0% on MATH, and verification-based reranking further raised accuracy to 89.1% and 43.5% respectively.	Process supervision is itself generated automatically, so quality depends on the automatic data construction pipeline.	high	abstract-only	Shows AI-generated step-level reward signals can drive LLM self-improvement without human annotations.
romeraparedes2024funsearch	benchmark	Pretrained LLM paired with a systematic evaluator (FunSearch) on the cap set problem and online bin packing	quality of discovered constructions/heuristics vs best-known results	FunSearch, an evolutionary procedure pairing a pretrained LLM with a systematic evaluator, discovered new constructions of large cap sets going beyond the best-known ones in both finite-dimensional and asymptotic cases and found new online bin packing heuristics that improve on widely used baselines.	LLMs can confabulate, so the method depends on the external evaluator to filter invalid programs; abstract reports no numeric margins.	high	abstract-only	LLM-plus-evaluator search yields novel mathematical discoveries beyond prior human knowledge, evidence for RSI in open problems.
silver2018alphazero	benchmark	AlphaZero (deep CNN + MCTS), trained 700k steps from random initialization in chess, shogi, Go; evaluated vs Stockfish, Elmo, AlphaGo Zero in 100-game matches	Win rate and Elo in 100-game matches at 1 min/move	AlphaZero, trained tabula rasa purely by self-play, scored 28 wins and 72 draws with 0 losses vs Stockfish and 90 wins vs Elmo in 100-game matches, surpassing the world-champion programs within 4 hours (chess, ~300k steps) and under 2 hours (shogi, ~110k steps) while searching only ~80k positions/sec vs Stockfish's 70M.	Enormous compute (5,000 first-gen TPUs for self-play, 64 for training); restricted to deterministic perfect-information games; some later replications questioned the exact hardware/timing claims.	high	full-text	Canonical positive result: self-play with no human data produces superhuman capability — the historical template for RSI feasibility arguments.
wang2023wizardlm	benchmark	LLaMA-7B fine-tuned on 250K Evol-Instruct instructions (4 evolution epochs over Alpaca's 52K seeds, generated via ChatGPT); Evol-Instruct testset (218 items, 29 skills) and Vicuna testset	human win rates; GPT-4 automatic evaluation scores	Evol-Instruct-trained WizardLM beat Vicuna-7B by 12.4% (Evol-Instruct testset) and 3.8% (Vicuna testset) win rate in human evaluation, and beat ChatGPT 42.9% vs 35.0% on high-difficulty (level >= 8) instructions while reaching >90% of ChatGPT's GPT-4 score on 17 of 29 skills.	GPT-4 automatic evaluation conflicted with human evaluation on hard skills; testset may not cover all domains; authors flag scalability/reliability limits of LLM-based evaluation; evolved instructions could be harmful or misleading.	high	full-text	Demonstrates LLM-driven instruction evolution (AI rewriting its own training data into harder tasks) improves capability, a key synthetic-data self-evolution mechanism for RSI.
rie2022modern	benchmark	Self-referential weight matrix (SRWM) using outer products and the delta update rule, evaluated on Omniglot and Mini-ImageNet few-shot classification (Conv-4-32 backend), sequential multi-task few-shot adaptation, and multi-task RL on 6 ProcGen environments (Bigfish, Fruitbot, Maze, Leaper, Plunder, Starpilot; easy distribution, 300M steps, IMPALA/Torchbeast, 3 seeds x 3 test splits)	Few-shot test accuracy; ProcGen normalized aggregated scores (x100)	SRWM matched strong baselines in few-shot learning (Omniglot 1-shot 97.4%, Mini-ImageNet 1-shot 47.0% and 5-shot 61.4%, comparable to DeltaNet 97.2/47.0/62.7 and SNAIL 99.1/45.1/55.2) and in multi-task ProcGen RL achieved the best test score of 20.0 vs DeltaNet 18.6, feed-forward 16.4, LSTM 15.7, and 15.3 for 'Fake SR' (SRWM without the self-modification mechanism), showing self-modification itself contributes to adaptation.	Gains over the DeltaNet baseline are small; self-modification is entangled with memory; experiments use small-scale architectures and tasks that do not explicitly require memory.	moderate	full-text	Practical existence proof of in-weight recursive self-modification: a scalable neural network that learns to modify its own weight matrix at runtime, relevant to the weight-level branch of RSI.
schmidhuber2007goedel	theoretical	n/a (formal framework: a Godel machine whose initial code contains a problem solver, an axiomatic system A describing hardware/software/environment/utility, and an O()-optimal proof searcher that rewrites any part of the code once a proof of usefulness is found; universal or space-bounded Turing machine hardware)	n/a (proofs only; no experiments or implementations)	No empirical numbers - the paper proves the Global Optimality Theorem (4.1): any self-rewrite the embedded proof searcher proves useful is globally optimal with no local maxima, and Theorem 5.1 gives an O()-optimal initial proof search (BIOPS), while Godelian limits mean provably useful rewrites that cannot be proven must be ignored.	Purely theoretical; proof search is computationally intractable in practice; no implementation or empirical demonstration exists.	high	full-text	The canonical theoretical model of provably optimal self-improvement (and its Godelian impossibility limits) - foundational for the formal side of the RSI review.

Swipe sideways to see all columns.

References

Nivel, E. et al. (2013). Bounded Recursive Self-Improvement — arXiv preprint. Full text read. Early (2013) conceptual architecture for bounded RSI via self-modeling and reflection; useful as historical/conceptual grounding but carries no quantitative evidence of gains.doi:10.48550/arxiv.1312.6764
Yampolskiy, Roman V. (2015). From Seed AI to Technological Singularity via Recursively Self-Improving Software — arXiv preprint. Full text read. Foundational conceptual reference for RSI (definitions, taxonomy, convergence theory, security) widely cited in the RSI literature.doi:10.48550/arxiv.1502.06512
Yampolskiy, Roman V. (2015). On the Limits of Recursively Self-Improving AGI — Lecture notes in computer science. Abstract only. Foundational theoretical framing of computational limits on recursive self-improvement.doi:10.1007/978-3-319-21365-1_40
Wang, Wenyi (2018). A Formulation of Recursive Self-Improvement and Its Possible Efficiency — arXiv preprint. Full text read. Formal existence argument that efficient (logarithmic-time) recursive self-improvement is computable in a restricted Markov setting, grounding theoretical feasibility claims for RSI.doi:10.48550/arxiv.1805.06610
Steunebrink, Bas R. et al. (2016). Growing Recursive Self-Improvers — Lecture notes in computer science. Abstract only. Conceptual contribution on formalizing when an agent can safely and effectively modify itself.doi:10.1007/978-3-319-41649-6_13
Feferman, Solomon (2006). Are There Absolutely Unsolvable Problems? Godel's Dichotomy — Philosophia Mathematica. Abstract only. Informs the debate on whether recursively self-improving machines face hard limits relative to human mathematical capability.doi:10.1093/philmat/nkj003
Muehlhauser, Luke & Salamon, Anna (2012). Intelligence Explosion: Evidence and Import — The frontiers collection. Abstract only. Surveys the probability case for an intelligence explosion, a central premise of RSI risk analysis.doi:10.1007/978-3-642-32560-1_2
Chalmers, David J. (2016). The Singularity — Science Fiction and Philosophy (Wiley). Abstract only. Canonical philosophical statement of the recursive self-improvement explosion scenario.doi:10.1002/9781118922590.ch16
Jebari, Karim & Lundborg, Joakim (2018). The intelligence explosion revisited — foresight. Abstract only. Key counterpoint to explosive RSI, arguing gradual transitions keep AI controllable.doi:10.1108/fs-04-2018-0042
Whitfill, Parker & Wu, Cheryl (2025). Will Compute Bottlenecks Prevent an Intelligence Explosion? — arXiv preprint. Full text read. First empirical estimate of compute-labor substitutability at frontier labs; whether compute bottlenecks prevent an intelligence explosion hinges entirely on which specification is correct.doi:10.48550/arxiv.2507.23181
MacAskill, William & Moorhouse, Fin (2025). Preparing for the Intelligence Explosion — arXiv preprint. Full text read. Frames AGI preparedness beyond alignment: if RSI leads to an intelligence explosion, governance and decision-making challenges arrive in rapid succession.doi:10.48550/arxiv.2506.14863
Evans, James et al. (2026). Agentic AI and the next intelligence explosion — arXiv preprint. Abstract only. Reframes RSI as collective/social scaling of intelligence and proposes institutional alignment as the control mechanism.doi:10.48550/arxiv.2603.20639
Field, Severin et al. (2026). AI Researchers' Views on Automating AI R&D and Intelligence Explosions — arXiv preprint. Full text read. Expert-opinion evidence that recursive self-improvement / AI-automated R&D is treated as a severe, urgent risk with contested timelines and governance — grounds the RSI risk debate in practitioner beliefs.doi:10.48550/arxiv.2603.03338
Falahati, Ali et al. (2025). The Alignment Game: A Theory of Long-Horizon Alignment Through Recursive Curation — arXiv preprint. Full text read. First formal treatment of long-horizon alignment under recursive retraining — shows self-consuming/self-improving loops have structural failure modes (collapse, power asymmetry, path dependence).doi:10.48550/arxiv.2511.12804
Mobahi, Hossein et al. (2020). Self-Distillation Amplifies Regularization in Hilbert Space — arXiv (Cornell University). Full text read. Foundational theoretical result explaining why iterative self-training (self-distillation) yields non-monotonic returns and eventually collapses - a key formal limit for recursive self-improvement loops.doi:10.48550/arxiv.2002.05715
Frei, Spencer et al. (2021). Self-training Converts Weak Learners to Strong Learners in Mixture Models — arXiv preprint. Full text read. Foundational theory that self-training converts weak learners to strong learners using only unlabeled data — a mathematical basis for self-improvement from self-generated data.doi:10.48550/arxiv.2106.13805
Zhang, Shuai et al. (2022). How does unlabeled data improve generalization in self-training? A one-hidden-layer theoretical analysis — arXiv preprint. Full text read. First theoretical characterization of iterative self-training on nonlinear networks — formal backing for the premise that training on one's own pseudo-labeled outputs provably improves generalization under suitable conditions.doi:10.48550/arxiv.2201.08514
Anthony, Thomas et al. (2017). Thinking Fast and Slow with Deep Learning and Tree Search — arXiv preprint. Full text read. Foundational expert-iteration / self-play algorithm - the template for self-improvement loops later applied to LLMs (cited by SPIN, CodeIt, Constitutional AI).doi:10.48550/arxiv.1705.08439
Silver, David et al. (2017). Mastering the game of Go without human knowledge — Nature. Abstract only. Canonical empirical demonstration of a recursive self-play improvement loop exceeding human-level performance.doi:10.1038/nature24270
Soemers, Dennis J. N. J. et al. (2020). Manipulating the Distributions of Experience used for Self-Play Learning in Expert Iteration — arXiv preprint. Full text read. Shows that how self-play experience is sampled/weighted inside the loop materially changes self-improvement dynamics — data curation matters in RSI.doi:10.48550/arxiv.2006.00283
Hernandez, Daniel et al. (2022). BRExIt: On Opponent Modelling in Expert Iteration — arXiv preprint. Full text read. Expert Iteration variant showing opponent modelling accelerates self-play learning, relevant to multi-agent and best-response dynamics in RSI.doi:10.48550/arxiv.2206.00113
Xie, Qizhe et al. (2020). Self-Training With Noisy Student Improves ImageNet Classification — 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Abstract only. Early large-scale evidence that iterated self-training (student-becomes-teacher) yields compounding gains, a precursor RSI mechanism.doi:10.1109/cvpr42600.2020.01070
Zelikman, Eric et al. (2022). STaR: Bootstrapping Reasoning With Reasoning — arXiv preprint. Full text read. The foundational self-training loop (generate rationale -> filter by correctness -> finetune -> repeat) that most later LLM self-improvement and RSI work builds on; shows a model improving itself from its own outputs.doi:10.48550/arxiv.2203.14465
Zelikman, Eric et al. (2024). Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking — arXiv preprint. Full text read. Generalizes self-improvement from curated QA to arbitrary text — the LM learns to generate useful internal reasoning at every token, evidence that self-generated thought improves the model itself at scale.doi:10.48550/arxiv.2403.09629
Hosseini, Arian et al. (2024). V-STaR: Training Verifiers for Self-Taught Reasoners — arXiv preprint. Full text read. Shows that exploiting both correct and incorrect self-generated solutions (via a DPO verifier) strengthens iterative self-improvement loops.doi:10.48550/arxiv.2402.06457
Zeng, Weihao et al. (2024). B-STaR: Monitoring and Balancing Exploration and Exploitation in Self-Taught Reasoners — arXiv preprint. Full text read. Explains why self-improvement saturates after 3-5 iterations (exploration collapse and weakening reward discrimination) and shows monitoring + dynamic balancing extends the improvement trajectory — key mechanistic evidence for RSI limits.doi:10.48550/arxiv.2412.17256
Huang, Jiaxin et al. (2023). Large Language Models Can Self-Improve — Proceedings of EMNLP 2023. Abstract only. Direct evidence that LLMs can improve their own reasoning from unlabeled data, a core RSI mechanism.doi:10.18653/v1/2023.emnlp-main.67
Wang, Yizhong et al. (2022). Self-Instruct: Aligning Language Models with Self-Generated Instructions — Annual Meeting of the Association for Computational Linguistics. Full text read. Foundational bootstrapping result showing a model can improve its own instruction-following using its own generations, a core synthetic-data self-improvement datapoint for RSI.doi:10.48550/arxiv.2212.10560
Weng, Yixuan et al. (2023). Large Language Models are Better Reasoners with Self-Verification — Findings of EMNLP 2023. Abstract only. Self-verification as a lightweight, weight-free self-improvement mechanism for LLM reasoning.doi:10.18653/v1/2023.findings-emnlp.167
Gandhi, K. et al. (2025). Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs — arXiv.org. Full text read. Identifies initial cognitive behaviors (verification, backtracking) as preconditions for effective RL self-improvement — explains plateauing and shows how to engineer self-improving models via data curation.doi:10.48550/arxiv.2503.01307
Sun, Yutao et al. (2025). The Self-Improvement Paradox: Can Language Models Bootstrap Reasoning Capabilities without External Scaffolding? — arXiv preprint. Full text read. Affirmative evidence for genuine self-improvement with zero external supervision (no seed data, no third-party models) in math reasoning — a key feasibility data point for RSI.doi:10.48550/arxiv.2502.13441
Shafayat, Sheikh et al. (2025). Can Large Reasoning Models Self-Train? — arXiv preprint. Full text read. Central evidence for both the promise and the hard limit of RSI: self-rewarded RL improves capability and label quality short-term, then collapses completely - making feedback design the key challenge.doi:10.48550/arxiv.2505.21444
Butt, Natasha et al. (2024). CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay — arXiv preprint. Full text read. Example of self-improvement via expert iteration with hindsight relabeling and prioritized replay in sparse-reward domains - bounded self-refinement in a neuro-symbolic loop.doi:10.48550/arxiv.2402.04858
Zhao, Zirui et al. (2024). Automatic Curriculum Expert Iteration for Reliable LLM Reasoning — arXiv preprint. Full text read. Shows self-improvement can be coupled with calibrated abstention — the model learns both to extend its reasoning and to refuse beyond its limits, relevant to safe self-evolving systems.doi:10.48550/arxiv.2410.07627
He, Yifei et al. (2024). Semi-Supervised Reward Modeling via Iterative Self-Training — arXiv preprint. Full text read. Self-training on unlabeled data reduces human-labeling dependency in reward modeling, automating a key RLHF ingredient relevant to RSI.doi:10.48550/arxiv.2409.06903
Wu, Mingqi et al. (2026). Why Self-Training Helps and Hurts: Denoising vs. Signal Forgetting — arXiv preprint. Full text read. Formalizes the denoising-vs-signal-forgetting tradeoff that dictates how many self-improvement iterations help before degradation - a limit result for iterative self-training.doi:10.48550/arxiv.2602.14029
Jang, Hyosoon et al. (2025). Self-Training Large Language Models with Confident Reasoning — arXiv preprint. Full text read. Shows the self-training loop's outcome hinges on the quality of the self-generated training signal: reasoning-level (not answer-level) confidence is what makes self-improvement work.doi:10.48550/arxiv.2505.17454
Luo, Ne et al. (2025). Self-Training Large Language Models for Tool-Use Without Demonstrations — arXiv preprint. Full text read. Nuanced counter-evidence: self-synthesized training data helps selectively (where the model's own knowledge is insufficient) - self-improvement via self-generated traces is not uniform.doi:10.48550/arxiv.2502.05867
Yuan, Siyu et al. (2025). Agent-R: Training Language Model Agents to Reflect via Iterative Self-Training — arXiv preprint. Full text read. Learns error correction from the model's own failed trajectories via MCTS splicing - iterative self-improvement of agentic recovery without expert critique data.doi:10.48550/arxiv.2501.11425
Madaan, Aman et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback — arXiv (Cornell University). Full text read. Canonical evidence that a single frozen model can iteratively improve its own outputs at inference time - the baseline for intrinsic self-improvement.doi:10.48550/arxiv.2303.17651
Du, Yilun et al. (2023). Improving Factuality and Reasoning in Language Models through Multiagent Debate — International Conference on Machine Learning. Full text read. Self-correction via multi-agent critique improves reasoning and factuality with no external supervision — a society-of-minds route to self-improvement and synthetic training data.doi:10.48550/arxiv.2305.14325
Yao, Shunyu et al. (2023). Tree of Thoughts: Deliberate Problem Solving with Large Language Models — arXiv (Cornell University). Full text read. Established self-evaluation plus search as an inference-time self-improvement mechanism for LLMs, later reused in self-play, self-correction, and self-evolving agent pipelines.doi:10.48550/arxiv.2305.10601
Pan, Liangming et al. (2023). Automatically Correcting Large Language Models: Surveying the landscape of diverse self-correction strategies — arXiv (Cornell University). Full text read. Provides the standard taxonomy (feedback source, format, timing, refinement strategy) that frames the self-correction literature feeding into the RSI review.doi:10.48550/arxiv.2308.03188
Liu, Dancheng et al. (2024). Large Language Models have Intrinsic Self-Correction Ability — arXiv.org. Full text read. Argues intrinsic self-correction is real when prompts are fair and sampling is deterministic - conditions under which self-feedback loops yield net gains.doi:10.48550/arxiv.2406.15673
Li, Loka et al. (2024). Confidence Matters: Revisiting Intrinsic Self-Correction Capabilities of Large Language Models — arXiv.org. Full text read. Shows the same self-feedback loop can help or hurt depending on gating - confidence-aware self-correction is a reliability condition for self-improvement.doi:10.48550/arxiv.2402.12563
Tsui, Ken (2025). Self-Correction Bench: Uncovering and Addressing the Self-Correction Blind Spot in Large Language Models — arXiv preprint. Full text read. Shows self-correction failure is an activation problem caused by training-data composition, not a capability gap — critical for whether RSI loops can catch their own errors, with a cheap fix (correction traces).doi:10.48550/arxiv.2507.02778
Sanz-Guerrero, Mario & Wense, K. (2025). Corrective In-Context Learning: Evaluating Self-Correction in Large Language Models — The Sixth Workshop on Insights from Negative Results in NLP. Full text read. Negative result: injecting the model's own errors plus corrections into context confuses rather than improves it - a caution for RSI designs that feed mistakes back into the prompt.doi:10.48550/arxiv.2503.16022
Moskvoretskii, Viktor et al. (2025). Self-Taught Self-Correction for Small Language Models — arXiv.org. Full text read. Shows weight-update self-improvement on self-generated corrections works for small models without external feedback, but gains are modest and filter selectivity is the critical design lever.doi:10.48550/arxiv.2503.08681
Chen, Kuan-Yen et al. (2026). The Self-Correction Illusion: Role Relabeling Gates Explicit Error Flagging in Large Language Models — arXiv preprint. Full text read. Challenges the 'LLMs cannot self-correct' narrative by showing the failure is partly a chat-template role-label artifact - a confound for any self-correction-based RSI evidence.doi:10.48550/arxiv.2606.05976
Rahmani, Hossein A. et al. (2025). Self-Correcting Large Language Models: Generation vs. Multiple Choice — arXiv preprint. Full text read. Evidence that iterative self-correction yields diminishing returns and format-dependent failure modes (drift vs inertia) - a ceiling result relevant to RSI loop design.doi:10.48550/arxiv.2511.09381
Song, Xiaoshuai et al. (2025). ProgCo: Program Helps Self-Correction of Large Language Models — Annual Meeting of the Association for Computational Linguistics. Full text read. Evidence that intrinsic self-correction helps only when verification is strong (program-driven): naive self-correction frequently hurts, a key caveat for RSI loops relying on self-feedback.doi:10.48550/arxiv.2501.01264
Zhang, Che et al. (2024). Learning to Check: Unleashing Potentials for Self-Correction in Large Language Models — arXiv.org. Full text read. Shows intrinsic self-correction fails unless the model is explicitly trained to check — evidence that self-improvement loops need learned verifiers rather than naive prompting, a central constraint for RSI.doi:10.48550/arxiv.2402.13035
Ganguli, Deep et al. (2023). The Capacity for Moral Self-Correction in Large Language Models — arXiv (Cornell University). Full text read. Evidence that self-correction capability scales with model size and RLHF — relevant both to safety of self-improving systems and to the claim that bigger models can steer their own behavior.doi:10.48550/arxiv.2302.07459
Yuan, Weizhe et al. (2024). Self-Rewarding Language Models — arXiv preprint. Full text read. Canonical demonstration that an LLM can serve as its own reward model and improve both instruction-following and self-judging ability across iterations - the archetypal self-rewarding RSI loop.doi:10.48550/arxiv.2401.10020
Wu, Tianhao et al. (2024). Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge — Conference on Empirical Methods in Natural Language Processing. Full text read. Shows that improving the judge along with the actor sustains iterative self-improvement where actor-only self-rewarding saturates - a key mechanism result for RSI.doi:10.48550/arxiv.2407.19594
Zhou, Xin et al. (2025). Self-Consistency of the Internal Reward Models Improves Self-Rewarding Language Models — arXiv.org. Full text read. Diagnoses unreliability in self-generated preference data (a core ingredient of self-improvement loops) and shows consistency constraints on internal rewards substantially improve self-rewarding training.doi:10.48550/arxiv.2502.08922
Wang, Yidong et al. (2025). Temporal Self-Rewarding Language Models: Decoupling Chosen-Rejected via Past-Future — arXiv preprint. Full text read. Shows iterative self-rewarding loops can degrade as chosen/rejected responses converge, and temporal decoupling sustains learning - a self-correction loop stability datapoint for RSI.doi:10.48550/arxiv.2508.06026
Tan, Chuyi et al. (2025). Breaking the Self-Confirming Loop: Diagnosing and Mitigating Systemic Reward Bias in Self-Rewarding RL — arXiv preprint. Full text read. Diagnoses why self-rewarding RL is unstable (confidence-coupled over-reward creates a self-confirming loop) and provides an ensemble-based fix — central to RSI stability.doi:10.48550/arxiv.2510.08977
Fu, Shi et al. (2026). Why Self-Rewarding Works: Theoretical Guarantees for Iterative Alignment of Language Models — arXiv preprint. Full text read. Provides the first theoretical explanation for why self-rewarding self-improvement converges despite poor initialization — supports stability of iterative self-alignment loops.doi:10.48550/arxiv.2601.22513
Xu, Ruijie et al. (2024). Just Say What You Want: Only-prompting Self-rewarding Online Preference Optimization — arXiv preprint. Full text read. Demonstrates a discriminator-free self-rewarding loop for smaller models - self-improvement without external reward models or GPT-4 API feedback.doi:10.48550/arxiv.2409.17534
Lee, Harrison et al. (2023). RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback — arXiv (Cornell University). Full text read. Foundational evidence that models can improve using AI (including their own) feedback instead of human labels - scalable self-improvement via RLAIF/d-RLAIF.doi:10.48550/arxiv.2309.00267
Sun, Zhiqing et al. (2023). Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision — arXiv (Cornell University). Full text read. Self-alignment from scratch: the model generates its own training data under human-written principles with minimal supervision — an alignment-oriented precursor of RSI.doi:10.48550/arxiv.2305.03047
Singh, Prof. Satya & Sharma, Ratnesh (2024). Reinforcement Learning from AI Feedback A Review — International Journal of Scientific Research in Computer Science Engineering and Information Technology. Abstract only. Automating the feedback signal (AI feedback) removes the human bottleneck from the alignment loop relevant to self-improvement.doi:10.32628/cseit24104135
Yin, Shangjian et al. (2025). Aligning Large Language Models via Fully Self-Synthetic Data — Annual Meeting of the Association for Computational Linguistics. Full text read. Shows a model can align itself with fully self-generated prompts, responses, and preferences - a closed-loop, human-data-free self-improvement recipe.doi:10.48550/arxiv.2510.06652
Chen, Zixiang et al. (2024). Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models — International Conference on Machine Learning. Full text read. Direct positive evidence that self-play can convert a weak LLM into a stronger one without new human data - a central result for LLM self-improvement.doi:10.48550/arxiv.2401.01335
Wu, Yue et al. (2024). Self-Play Preference Optimization for Language Model Alignment — arXiv preprint. Full text read. Game-theoretic self-play alignment with provable Nash convergence, showing iterative self-play improves LLMs without external labels - theory plus strong empirics for RSI.doi:10.48550/arxiv.2405.00675
Dong, Guanting et al. (2024). Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models — International Conference on Learning Representations. Full text read. Shows LLMs can synthesize their own verifiable training data and self-improve without human annotation — a concrete closed-loop self-improvement mechanism with execution feedback.doi:10.48550/arxiv.2406.13542
Xiao, Yao et al. (2025). Revisiting Self-Play Preference Optimization: On the Role of Prompt Difficulty — arXiv preprint. Full text read. Shows self-play preference optimization gains are prompt-difficulty-dependent - naive scaling of self-play data can yield diminishing or negative returns.doi:10.48550/arxiv.2510.05534
Wang, Yibo et al. (2025). SPACE: Noise Contrastive Estimation Stabilizes Self-Play Fine-Tuning for Large Language Models — Neural Information Processing Systems. Full text read. Diagnoses why self-play fine-tuning with gap objectives is unstable and proposes a provably convergent alternative, evidence on limits and stabilization of self-play RSI.doi:10.48550/arxiv.2512.07175
Tang, Xiaohang et al. (2025). RSPO: Regularized Self-Play Alignment of Large Language Models — arXiv preprint. Full text read. Shows regularization to a reference policy is crucial in self-play alignment — iterative self-training without it over-optimizes; provides convergence guarantees for the regularized game.doi:10.48550/arxiv.2503.00030
Jana, Swadesh et al. (2026). GASP: Guided Asymmetric Self-Play For Coding LLMs — arXiv preprint. Full text read. Goalpost-grounded asymmetric self-play provides a curriculum mechanism for RSI where a teacher generates problems at the student's frontier, with verified gains over unguided self-play.doi:10.48550/arxiv.2603.15957
Wilf, Alex et al. (2025). Propose, Solve, Verify: Self-Play Through Formal Verification — arXiv preprint. Full text read. Positive RSI datapoint: self-play without any human data works in code generation when the correctness signal is sound, contrasting with brittle unit-test-based self-play.doi:10.48550/arxiv.2512.18160
Dong, Kefan & Ma, Tengyu (2025). STP: Self-play LLM Theorem Provers with Iterative Conjecturing and Proving — arXiv preprint. Full text read. Direct evidence for self-play-based recursive improvement: conjecturer-prover loop keeps improving without new human data — a key RSI mechanism demonstration.doi:10.48550/arxiv.2502.00212
Liu, Mickel et al. (2025). Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models — arXiv.org. Full text read. Self-play attacker-defender co-evolution as an RSI loop that improves safety without external red teams - evidence for adversarial self-improvement.doi:10.48550/arxiv.2506.07468
Fu, Yao et al. (2023). Improving Language Model Negotiation with Self-Play and In-Context Learning from AI Feedback — arXiv.org. Full text read. Early evidence of autonomous mutual improvement of LLMs via self-play plus AI feedback (AlphaGo-Zero-style loop), and that a capability/alignment threshold gates who can self-improve.doi:10.48550/arxiv.2305.10142
Chojecki, Przemyslaw (2025). Self-Improving AI Agents through Self-Play — arXiv preprint. Full text read. Formalizes the verifier-noise condition for stable self-improvement - theoretical backing for the practitioner maxim 'strengthen the verifier, not the generator'.doi:10.48550/arxiv.2512.02731
Shumailov, Ilia et al. (2024). AI models collapse when trained on recursively generated data — Nature. Abstract only. Key caution for recursive training loops: without fresh human data, self-consumption degrades models irreversibly.doi:10.1038/s41586-024-07566-y
Alemohammad, Sina et al. (2023). Self-Consuming Generative Models Go MAD — arXiv (Cornell University). Full text read. Foundational negative result on model collapse in self-consuming training loops - the key caution for RSI via synthetic-data bootstrapping.doi:10.48550/arxiv.2307.01850
Gerstgrasser, Matthias et al. (2024). Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data — arXiv (Cornell University). Full text read. Key evidence that recursive training on self-generated data need not degrade performance if data accumulates, tempering model-collapse pessimism in the RSI debate.doi:10.48550/arxiv.2404.01413
Fu, Shi Hong et al. (2024). Towards Theoretical Understandings of Self-Consuming Generative Models — arXiv (Cornell University). Full text read. Formalizes model collapse in self-consuming loops and gives conditions (real-data mixing, sample scaling) under which recursive training on own outputs stays stable — directly relevant to RSI feasibility and collapse risk.doi:10.48550/arxiv.2402.11778
Ferbach, Damien et al. (2024). Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences — Advances in Neural Information Processing Systems 37. Abstract only. Formal result that human/curation feedback in the retraining loop can prevent collapse and steer self-improvement, but amplifies reward biases.doi:10.52202/079017-3256
Gillman, Nate et al. (2024). Self-Correcting Self-Consuming Loops for Generative Model Training — arXiv preprint. Full text read. Shows a concrete mechanism (self-correction of synthetic data) that stabilizes recursive generative training, relevant to preventing collapse in closed-loop RSI.doi:10.48550/arxiv.2402.07087
Kazdan, Joshua et al. (2024). Collapse or Thrive? Perils and Promises of Synthetic Data in a Self-Generating World — arXiv (Cornell University). Full text read. Central evidence for the model-collapse debate: whether recursive self-generation degrades models depends critically on the data workflow (replace vs accumulate), shaping RSI feasibility forecasts.doi:10.48550/arxiv.2410.16713
Zhao, Hongru et al. (2025). Convergence and Stability Analysis of Self-Consuming Generative Models with Heterogeneous Human Curation — arXiv preprint. Full text read. Theoretical account of self-consuming retraining (training on one's own outputs): identifies when recursive retraining converges, and shows reference anchoring is what buys stability — directly relevant to model-collapse debates in RSI.doi:10.48550/arxiv.2511.09002
Wei, Xiukun & Zhang, Xueru (2025). Self-Consuming Generative Models with Adversarially Curated Data — arXiv preprint. Full text read. Extends self-consuming-loop theory to noisy/adversarial curation: retraining on self-generated, user-curated data is vulnerable to poisoning, bounding the robustness of preference-driven RSI.doi:10.48550/arxiv.2505.09768
Wyllie, Sierra et al. (2024). Fairness Feedback Loops: Training on Synthetic Data Amplifies Bias — ACM Conference on Fairness, Accountability, and Transparency (FAccT 2024). Abstract only. Documents how recursive training loops encode prior mistakes and biases, and a possible corrective intervention.doi:10.1145/3630106.3659029
Kim, Jeonghye et al. (2026). Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs? — arXiv preprint. Full text read. Documents a failure mode of self-improvement: distilling confident, privileged-information teacher traces can quietly destroy generalization - critical caution for RSI loops that train on their own confident outputs.doi:10.48550/arxiv.2603.24472
Wang, Meng et al. (2026). Denser $\neq$ Better: Limits of On-Policy Self-Distillation for Continual Post-Training — arXiv preprint. Full text read. Direct negative evidence that a self-distillation loop (model teaching itself token-by-token) is a fragile, collapse-prone RSI signal in continual settings.doi:10.48550/arxiv.2607.01763
Burns, Collin et al. (2023). Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision — arXiv preprint. Full text read. Quantifies the ceiling of weak-to-strong (self-)supervision: RSI loops inherit supervisor quality limits, and naive self-supervision recovers only part of the capability gap.doi:10.48550/arxiv.2312.09390
Lang, Hunter et al. (2024). Theoretical Analysis of Weak-to-Strong Generalization — arXiv preprint. Full text read. Theoretical foundation for why strong models improve on weak teachers' labels - underpins weak-to-strong supervision as an RSI-adjacent mechanism.doi:10.48550/arxiv.2405.16043
Moniri, Behrad & Hassani, Hamed (2025). On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspective — arXiv preprint. Full text read. Supplies theoretical grounding for weak-to-strong self-training: explains when a stronger student can exceed a weaker teacher's labels and the regularization conditions required, with implications for RSI via synthetic data.doi:10.48550/arxiv.2505.18346
Xue, Yihao et al. (2025). Representations Shape Weak-to-Strong Generalization: Theoretical Insights and Empirical Predictions — arXiv preprint. Full text read. Provides a measurable, label-free predictor of when weak supervision of a stronger model generalizes - directly relevant to superalignment-style RSI.doi:10.48550/arxiv.2502.00620
Somerstep, Seamus et al. (2025). Limitations of refinement methods for weak to strong generalization — arXiv preprint. Full text read. Theoretical ceiling on weak-to-strong / self-training approaches: self-generated labels cannot be refined to oracle quality, bounding what RSI via self-labeling can achieve.doi:10.48550/arxiv.2508.17018
Yang, Wenkai et al. (2024). Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization — arXiv preprint. Full text read. Cautions that weak-to-strong generalization (basis of superalignment-style RSI) can conceal deception by the stronger model exactly where the weak supervisor is blind.doi:10.48550/arxiv.2406.11431
Shin, Changho et al. (2024). Weak-to-Strong Generalization Through the Data-Centric Lens — arXiv preprint. Full text read. Data-centric account of when weak supervision (self-training/pseudo-labeling) can lift a strong model beyond its teacher — conditions under which self-improvement loops can escape the weak-teacher ceiling.doi:10.48550/arxiv.2412.03881
Geng, Scott et al. (2026). Weak-to-Strong Generalization is Nearly Inevitable (in Linear Models) — arXiv preprint. Full text read. Theoretical grounding that weak-to-strong generalization (a student surpassing a weaker teacher and itself) is 'almost inevitable', lending formal support to the feasibility of RSI loops.doi:10.48550/arxiv.2605.05742
Gao, Huan-ang et al. (2025). A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence — arXiv preprint. Full text read. Maps the self-evolving-agents landscape (mechanisms, stages, evaluation) that the RSI review can use to situate primary evidence.doi:10.48550/arxiv.2507.21046
Fang, Jinyuan et al. (2025). A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems — arXiv preprint. Full text read. Provides the conceptual scaffolding for the RSI review: a component-level map of where and how agents self-evolve, plus explicit safety considerations.doi:10.48550/arxiv.2508.07407
Zhou, Wangchunshu et al. (2024). Symbolic Learning Enables Self-Evolving Agents — arXiv preprint. Full text read. A systematic framework for self-evolving agents (agents as symbolic networks optimized with language back-propagation/gradient descent) — evidence that agent-level self-modification improves task performance data-centrically.doi:10.48550/arxiv.2406.18532
Sheng, Alex (2024). From Language Models to Practical Self-Improving Computer Agents — arXiv (Cornell University). Full text read. Early proof-of-concept that LLM agents can self-improve by writing their own scaffolding/tools; evidence is anecdotal but influential for agent-level RSI.doi:10.48550/arxiv.2404.11964
Zelikman, Eric et al. (2023). Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation — arXiv preprint. Full text read. First demonstration that an LLM can write code that calls itself to improve itself (recursive scaffolding improvement), plus early quantification of reward hacking / sandbox-bypass frequency in self-improving code.doi:10.48550/arxiv.2310.02304
Ishibashi, Yoichi et al. (2024). Can Large Language Models Invent Algorithms to Improve Themselves?: Algorithm Discovery for Recursive Self-Improvement through Reinforcement Learning — arXiv preprint. Full text read. Evidence for recursive self-improvement of the improvement algorithm itself (LLM autonomously discovering its own merging/fine-tuning methods).doi:10.48550/arxiv.2410.15639
Lu, Shuqi et al. (2026). AREX: Towards a Recursively Self-Improving Agent for Deep Research — arXiv preprint. Full text read. Directly an RSI paper: recursive self-improvement of answers via verification-driven refinement loops, evidence that RSI-style agents beat static search.doi:10.48550/arxiv.2607.21461
Wang, Zefeng et al. (2026). MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution — arXiv preprint. Full text read. A concrete bounded-recursion RSI implementation - evolving the improvement operator itself - with positive but modest, benchmark-dependent gains.doi:10.48550/arxiv.2607.05297
Nguyen, Michael et al. (2026). Recursive Self-Evolving Agents via Held-Out Selection — arXiv preprint. Full text read. Key evidence that the selection gate - not the sophistication of the evolved artifact - determines whether recursive context self-evolution is monotone-safe and free of catastrophic regression.doi:10.48550/arxiv.2606.28374
Jin, Yuxin et al. (2026). EXG: Self-Evolving Agents with Experience Graphs — arXiv preprint. Full text read. Evidence that non-parametric experience accumulation (graph-structured memory) lets agents improve during deployment without weight updates - a lightweight RSI mechanism.doi:10.48550/arxiv.2605.17721
Zhang, Xing et al. (2026). Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents — arXiv preprint. Full text read. Tackles the 'who grades the grader' problem in self-evolving loops: evaluation metrics can themselves be evolved, but anchor discipline (not lifecycle management) is the load-bearing safety ingredient — a Goodhart caution for any self-improving system.doi:10.48550/arxiv.2607.12790
Ying, Zonghao et al. (2026). SkillJack: Persistent Skill Backdoors in Self-Evolving Agents — arXiv preprint. Full text read. Reveals the experience-to-skill pipeline of self-evolving agents as a durable attack surface - self-improvement mechanisms can entrench malicious behaviors beyond source-record deletion.doi:10.48550/arxiv.2608.03509
Zhao, Weixiang et al. (2026). On Safety Risks in Experience-Driven Self-Evolving Agents — arXiv preprint. Full text read. Core safety evidence for the RSI review: self-evolving agents degrade on safety even from benign experience, exposing a fundamental safety-utility trade-off in experience-driven self-improvement.doi:10.48550/arxiv.2604.16968
Cai, Yuxuan et al. (2025). Building Self-Evolving Agents via Experience-Driven Lifelong Learning: A Framework and Benchmark — arXiv preprint. Full text read. Provides a measurement instrument (StuLife/StuGPA) for evaluating self-evolving agents - evidence that current LLMs are far from open-ended self-improvement.doi:10.48550/arxiv.2508.19005
Qian, Hongjin & Liu, Zheng (2025). MetaAgent: Toward Self-Evolving Agent via Tool Meta-Learning — arXiv preprint. Full text read. Strong demonstration that self-evolving agents (context + in-house tools, no parameter updates) can match or beat end-to-end RL-trained agents on knowledge discovery - evidence for context-level RSI.doi:10.48550/arxiv.2508.00271
Hebbar, Prannay et al. (2026). SIA: Self Improving AI with Harness & Weight Updates — arXiv preprint. Full text read. Concrete closed-loop self-improvement system updating both agent scaffold and model weights, a direct RSI implementation with consistent cross-domain gains.doi:10.48550/arxiv.2605.27276
Wang, Su et al. (2026). Phantom Guardrails: When Self-Improving Agent Harnesses Fix Failures That Never Happened — arXiv preprint. Full text read. Shows self-improving harness optimizers can hallucinate failures and add phantom guardrails invisible to suppression-only reward - a novel failure mode for RSI loops.doi:10.48550/arxiv.2607.13083
Yang, Zitong (2026). Continually self-improving AI — arXiv preprint. Full text read. A three-pronged program for continually self-improving AI: synthetic data for knowledge acquisition, self-bootstrapped pretraining, and AI-designed training algorithms.doi:10.48550/arxiv.2603.18073
Wang, Xuefei Julie et al. (2026). Knowledge-Centric Self-Improvement — arXiv preprint. Full text read. Argues RSI progress can be stored in a curated knowledge base rather than agent weights/prompts, a design-axis contribution with cost-efficiency evidence.doi:10.48550/arxiv.2607.19592
Aggarwal, Aditya & Ghalaty, Nahid Farhady (2026). Self-Improving AI Coding Agents Through Accumulated Behavioral Rules: A Closed-Loop Framework — arXiv preprint. Full text read. Evidence that bounded self-improvement via persistent external memory (accumulated rules from human review) works in production without weight updates; human review remains the signal source, so the loop is human-in-the-loop.doi:10.48550/arxiv.2607.13091
Yin, Xunjian et al. (2024). G\"odel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement — arXiv preprint. Full text read. A concrete LLM implementation of the Goedel-machine idea - the agent analyzes and modifies its own code to achieve recursive self-improvement.doi:10.48550/arxiv.2410.04444
Wu, Xuening et al. (2025). SGM: A Statistical Godel Machine for Risk-Controlled Recursive Self-Modification — arXiv (Cornell University). Full text read. First statistical safety layer for recursive self-modification - a risk-control gate that certifies edits before commit, foundational infrastructure for safe RSI loops.doi:10.48550/arxiv.2510.10232
Zhang, Jenny et al. (2025). Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents — arXiv preprint. Full text read. The strongest concrete demonstration of self-referential RSI in LLM systems: the agent edits its own codebase, improving both task performance and its capacity for further self-modification, with archive-based open-ended exploration and sandboxing.doi:10.48550/arxiv.2505.22954
Fang, Xinmin et al. (2026). Intelligence Flywheel: A Conjecture on Recursive Self-Improvement and the Dynamics of AI Progress — TechRxiv preprint. Abstract only. Contemporary conjecture quantifying AI's recursive contribution to its own development and arguing for recalibrated governance.doi:10.36227/techrxiv.176972273.31835965/v1
Meng, Fanqing et al. (2026). RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement — arXiv preprint. Full text read. Provides benchmark evidence of a 'discovery-reliability gap': current frontier agents can make useful data-centric discoveries for RSI but cannot yet convert feedback into consistent, monotone improvement.doi:10.48550/arxiv.2607.25886
Xue, Shuhan et al. (2026). PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents — arXiv preprint. Full text read. First benchmark isolating whether retained experience actually makes agents improve across sessions, with matched persistence on/off controls - measurement foundation for RSI in personal agents.doi:10.48550/arxiv.2608.04003
Sahoo, Subramanyam et al. (2026). SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement — arXiv preprint. Full text read. Makes alignment drift during RSI measurable (drift index, constraints, regression risk) and shows early improvement cycles are cheap while later gains carry rising alignment costs - direct evidence for the capability-alignment frontier.doi:10.48550/arxiv.2603.06333
Chen, Mingguang et al. (2026). Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops — arXiv preprint. Full text read. Provides the organizing taxonomy for the RSI review (bounded self-refinement vs open-ended RSI) and the verification-hierarchy lens for interpreting empirical self-improvement results.doi:10.48550/arxiv.2607.07663
Zhang, Jiang et al. (2026). Self-Reference in Large Language Models: The Introspection Threshold for Recursive Self-Improvement — arXiv preprint. Full text read. Provides the theoretical frame for when autonomous self-improvement is sustainable rather than degenerative, connecting RSI to self-reference, computability theory, and model collapse.doi:10.48550/arxiv.2607.04277
Hughes, Edward et al. (2024). Open-Endedness is Essential for Artificial Superhuman Intelligence — arXiv preprint. Full text read. Provides the conceptual framing linking open-endedness and self-improvement to ASI, useful for the review's theoretical framing of RSI.doi:10.48550/arxiv.2406.04268
Faldor, Maxence et al. (2024). OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code — arXiv preprint. Full text read. Demonstrates an open-ended self-improvement loop: models autonomously generate curricula (tasks + rewards) matched to agent ability — a step toward AI-Generating Algorithms and endless capability growth.doi:10.48550/arxiv.2405.15568
Zhang, Jenny et al. (2023). OMNI: Open-endedness via Models of human Notions of Interestingness — arXiv preprint. Full text read. Contributes auto-curriculum — an AI selecting its own next task to learn via a learned notion of interestingness — as a building block for self-improving AI and AI-Generating Algorithms.doi:10.48550/arxiv.2306.01711
Shinn, Noah et al. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning — NeurIPS 2023. Abstract only. Weight-free self-improvement via verbal feedback, showing LLM agents can learn from their own mistakes without fine-tuning.doi:10.52202/075280-0377
Bai, Yuntao et al. (2022). Constitutional AI: Harmlessness from AI Feedback — arXiv preprint. Full text read. Early large-scale demonstration of self-improvement via AI feedback replacing human labels - key evidence for scaling supervision in alignment.doi:10.48550/arxiv.2212.08073
Christiano, Paul et al. (2018). Supervising Strong Learners by Amplifying Weak Experts — arXiv preprint. Full text read. Foundational scalable-oversight proposal: build training signals for hard tasks by amplifying weak experts on easier subproblems - ancestor of recursive reward modeling in self-improvement supervision.doi:10.48550/arxiv.1810.08575
Leike, Jan et al. (2018). Scalable Agent Alignment via Reward Modeling: A Research Direction — arXiv preprint. Full text read. Seminal framework for recursive reward modeling - a canonical RSI architecture (agents training better agents via amplified evaluation).doi:10.48550/arxiv.1811.07871
Li, Xian et al. (2023). Self-Alignment with Instruction Backtranslation — arXiv preprint. Full text read. Demonstrates an iterative self-training loop (generate -> self-curate -> retrain) that improves instruction following without external distillation - direct RSI evidence.doi:10.48550/arxiv.2308.06259
Chen, Xinyun et al. (2023). Teaching Large Language Models to Self-Debug — arXiv preprint. Full text read. Evidence for bounded self-refinement in code generation via execution feedback and self-explanation - the 'code self-repair via execution feedback' category of RSI.doi:10.48550/arxiv.2304.05128
Guo, Shangmin et al. (2024). Direct Language Model Alignment from Online AI Feedback — arXiv preprint. Full text read. Evidence that on-policy (online) feedback - sampling from the current model - materially improves self-training alignment, a key design choice for RSI loops.doi:10.48550/arxiv.2402.04792
Huang, Jie et al. (2023). Large Language Models Cannot Self-Correct Reasoning Yet — arXiv preprint. Full text read. Critical negative result: intrinsic self-correction without external feedback fails, a key boundary condition for RSI claims about self-correction.doi:10.48550/arxiv.2310.01798
Gulcehre, Caglar et al. (2023). Reinforced Self-Training (ReST) for Language Modeling — arXiv preprint. Full text read. Canonical growing-batch self-training loop (sample-filter-finetune, repeat) demonstrating measurable iterative self-improvement in LLMs.doi:10.48550/arxiv.2308.08998
Tao, Zhengwei et al. (2024). A Survey on Self-Evolution of Large Language Models — arXiv preprint. Full text read. Provides the organizing taxonomy and open-problem list (autonomy levels, three-round improvement ceiling, model collapse, superalignment) that structures the RSI review.doi:10.48550/arxiv.2404.14387
Shumailov, Ilia et al. (2023). The Curse of Recursion: Training on Generated Data Makes Models Forget — arXiv preprint. Full text read. Canonical negative result for recursive self-improvement: training on self-generated data causes irreversible model collapse (tail loss), foundational for RSI risk analysis.doi:10.48550/arxiv.2305.17493
Wang, Guanzhi et al. (2023). Voyager: An Open-Ended Embodied Agent with Large Language Models — arXiv preprint. Full text read. Demonstrates compounding, compositional skill-library growth in a lifelong-learning embodied agent — positive agent-level evidence for RSI-like accumulation without parameter updates.doi:10.48550/arxiv.2305.16291
Olausson, Theo X. et al. (2023). Is Self-Repair a Silver Bullet for Code Generation? — arXiv preprint. Full text read. Central conditional evidence on self-repair for RSI: the model's own feedback is the bottleneck, so self-correction gains are limited unless feedback quality is externally boosted.doi:10.48550/arxiv.2306.09896
Novikov, Alexander et al. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery — arXiv preprint. Full text read. Evidence that LLM-driven evolution grounded in execution can improve critical infrastructure, including the training of the LLM underpinning AlphaEvolve itself - a partial self-referential closed loop.doi:10.48550/arxiv.2506.13131
Robeyns, Maxime et al. (2025). Self-Improving Coding Agent — arXiv preprint. Full text read. Direct empirical demonstration of scaffold-level recursive self-improvement in a coding agent, with clear saturation when the underlying model already performs well.doi:10.48550/arxiv.2504.15228
Wang, Xuezhi et al. (2022). Self-Consistency Improves Chain of Thought Reasoning in Language Models — arXiv preprint. Full text read. Shows a purely inference-time self-ensemble (no weight updates) yields large reasoning gains, a zero-parameter form of self-improvement relevant to RSI scaling.doi:10.48550/arxiv.2203.11171
Gou, Zhibin et al. (2023). CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing — arXiv preprint. Full text read. Distinguishes effective (tool-grounded) from ineffective (intrinsic) self-correction, establishing that external feedback is crucial for consistent LLM self-improvement.doi:10.48550/arxiv.2305.11738
Haluptzok, Patrick et al. (2024). Language Models Can Teach Themselves to Program Better — arXiv preprint. Full text read. Direct demonstration of an LM teaching itself to program better through self-play with machine verification - a canonical RSI result.doi:10.48550/arxiv.2207.14502
Zheng, Lianmin et al. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena — NeurIPS 2023. Abstract only. Enables scalable automated evaluation, a critical feedback component for closed-loop self-improvement pipelines.doi:10.52202/075280-2020
Gao, Leo et al. (2023). Scaling Laws for Reward Model Overoptimization — arXiv preprint. Full text read. Quantifies the Goodhart/overoptimization ceiling when optimizing against a learned proxy reward — a fundamental limit for self-improvement loops that optimize self-generated or proxy rewards.doi:10.48550/arxiv.2210.10760
Liu, Shengcai et al. (2024). Large Language Models as Optimizers — IEEE CEC 2024. Abstract only. Demonstrates LLMs optimizing their own instructions, a direct self-optimization loop at the prompt level.doi:10.1109/cec60901.2024.10611913
Lu, Chris et al. (2024). The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery — arXiv preprint. Full text read. Canonical demonstration of closed-loop AI research automation ('auto research' category in RSI); shows evaluator quality, not generation, is the bottleneck in automated discovery loops.doi:10.48550/arxiv.2408.06292
Wang, Peiyi et al. (2024). Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations — ACL 2024. Abstract only. Shows AI-generated step-level reward signals can drive LLM self-improvement without human annotations.doi:10.18653/v1/2024.acl-long.510
Romera-Paredes, Bernardino et al. (2023). FunSearch: Mathematical discoveries from program search with large language models — Nature. Abstract only. LLM-plus-evaluator search yields novel mathematical discoveries beyond prior human knowledge, evidence for RSI in open problems.doi:10.1038/s41586-023-06924-6
Silver, David et al. (2017). Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm — arXiv preprint. Full text read. Canonical positive result: self-play with no human data produces superhuman capability — the historical template for RSI feasibility arguments.doi:10.48550/arxiv.1712.01815
Xu, Can et al. (2023). WizardLM: Empowering large pre-trained language models to follow complex instructions — arXiv preprint. Full text read. Demonstrates LLM-driven instruction evolution (AI rewriting its own training data into harder tasks) improves capability, a key synthetic-data self-evolution mechanism for RSI.doi:10.48550/arxiv.2304.12244
Irie, Kazuki et al. (2022). A Modern Self-Referential Weight Matrix That Learns to Modify Itself — arXiv preprint. Full text read. Practical existence proof of in-weight recursive self-modification: a scalable neural network that learns to modify its own weight matrix at runtime, relevant to the weight-level branch of RSI.doi:10.48550/arxiv.2202.05780
Schmidhuber, Jürgen (2007). Gödel Machines: Fully Self-referential Optimal Universal Self-improvers — Cognitive Technologies. Full text read. The canonical theoretical model of provably optimal self-improvement (and its Godelian impossibility limits) - foundational for the formal side of the RSI review.doi:10.1007/978-3-540-68677-4_7