Pith.

Machine Learning · arXiv:2610.27610 v1 · submitted 20 Oct 2026 · 125 references

Auditing the Product, Not the Recursion: The Evaluation Evidence Behind Recursive Self-Improvement Claims

Zihan Chen, Di Zhu, Lei Nico Zheng

The paper says

“Recursive self-improvement (RSI) promises AI systems, such as self-evolving LLM agents, that become better at producing further improvements. Evaluating this promise requires separating product claims about the final system from process claims about the improvement process itself.”

more of the abstract

Existing surveys specify the evidence these claims require, but how often studies report that evidence remains unmeasured. We examine this question through full-text coding of 114 eligible studies in a reproducible arXiv search frame frozen in October 2026, including 90 that describe their own systems as RSI. The codebook links claims to evaluation designs. Independent human coding, third-reader adjudication, and a separate blind coding of improver comparisons assess coding reliability. Among the 90 RSI studies, 90% report product evaluation beyond development tasks under a broad criterion that includes partial or unstated separation, while 60% report evaluation on tasks unused for selection. Only 22% audit two or more reruns of the full loop. In a human-coded random sample, the RSI studies show the same contrast: 45% evaluate on tasks unused for selection and 14% audit two or more reruns. Of the 26 studies claiming a better improvement procedure, 23 report a related comparison, but only 4 compare earlier and later procedures from identical starting artifacts at a matched budget. Across these results and a comparison set of established works, evidence for repeatability and improved procedures remains less developed than product evaluation. Three propositions explain how selection effects, shared evaluator errors, and changes in system state limit what product evaluations can establish. Together, the findings support a claim-indexed reporting standard that connects each claim to the evaluation design needed to test it.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 21 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up. A well-executed, transparent meta-research study that credibly measures the gap between product and process evaluation in RSI papers. The central finding is robust, though the search-frame and coding-reliability caveats are substantive. This line is wrong

2 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 125 references say.

The paper says, at §9 (Limitations)

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

The paper acknowledges that the AI-assisted first pass erred substantially more often than the independent human coding (line 297) and that outside the double-coded sample 'individual codes have not been checked by a human' (line 576). The random-sample comparison shows directional differences: the census overstates selection-independent evaluation and understates scope gaps.

more of objection 1

The authors treat this transparently, but the central prevalence estimates (e.g., 60% for selection-independent evaluation, 22% for audited reruns) rest on a mix of adjudicated and non-adjudicated codes. The paper should explicitly state which estimates are based on the full census versus the adjudicated sample alone and, for the full-census estimates, provide the sensitivity analysis that would result if the non-adjudicated codes were adjusted to match the adjudicated shares for the measured items (e.g., 41% instead of 60% for selection-independent evaluation among strong-label studies). This would give readers a clear range rather than relying on the statement that the census 'may overstate product evaluation.'

The paper says, at §5.3 and Table 8

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.91):

“However, they were revised after the second coding of these same studies, so they do not show how accurate v1.1 is elsewhere.”

Pith says: Objection 2 of 2 · major

Table 8 compares the adjudicated double-coded sample with the other eligible studies on headline items. The Fisher p-values are reported, but the text (line 432) notes that the v1.1 codes of the sampled studies were 'revised after the second coding of these same studies, so they do not show how accurate v1.1 is elsewhere.'

more of objection 2

This is a critical caveat that undermines the value of the comparison between the adjudicated sample and the 'other eligible studies' column, because the v1.1 codes for the sampled studies are not representative of how v1.1 performs on unseen studies. The authors should either present the first-pass (pre-revision) codes for the sampled studies alongside the adjudicated codes to give an honest estimate of v1.1 accuracy, or remove the comparison and simply note that the adjudicated sample provides an estimate that does not depend on the AI-assisted pass.

6 minor objections

The paper says, at §3, Table 2

Pith says: minor objection 1 of 6

The definition of C3 (Repeatability) requires 'two or more independent reruns of the full loop, audited, with variation across runs and a stated target quantity (estimand).' However, the census item P3a records only whether at least two reruns have audit-level outcomes, not whether variation or an estimand is stated. The text should clarify this mismatch and note that the measured 22% is a lower bound on full compliance with C3.

The paper says, at §5.3, line 291

Pith says: minor objection 2 of 6

The agreement statistics (Cohen's κ between 0.81 and 0.91) for items like label, domain, and audited reruns are strong, but Table 16 shows that many evaluation items (A2, A4, A5, A6, P4) have κ below 0.40. The text should acknowledge that the high agreement holds only for a subset of items and that the codebook requires further refinement for others, as is already implied by the v1.1 rule changes.

show 4 more

The paper says, at §6.2, Table 5

Pith says: minor objection 3 of 6

The strict same-start rule (identical starting artifacts, not equal scores) is applied after the independent coding. The table and text should state clearly that this rule reclassified two comparisons that the independent coder had counted as same-start. Currently the description on lines 370-371 mentions this but the table itself does not indicate which entries were reclassified.

The paper says, at §8.2, Table 9

Pith says: minor objection 4 of 6

The reporting standard in Table 9 lists 'the share of strong-label studies, or of the studies making the claim, that report the strict indicator of Table 4.' However, the Census rate column gives percentages (e.g., 60% for C1, 60% for C2) but does not specify the denominator (all 90 studies or only claimants). Adding the denominator in parentheses (e.g., '60% (54/90)' and '60% of claimants (27/45)') would eliminate ambiguity.

The paper says, at §5.2, evidence tiers

Pith says: minor objection 5 of 6

The evidence tiers (E0–E4) are used in Section 6.4 and Table 6, but the definition of E4 requires 'an improver-related comparison at a matched budget (P4 = Y).' However, P4 = Y includes evolving-versus-frozen comparisons and other designs, not just same-start comparisons. This means E4 bundles designs that answer different questions, as acknowledged in the text. The paper should add a note in §5.2 that E4 is a summary indicator and that claim-specific analysis (Section 6.2) disaggregates it.

The paper says, at §9, line 570

Pith says: minor objection 6 of 6

The limitation that 'the codes describe what papers report, not all experiments their authors may have conducted' is well stated. However, the paper also notes that a study that runs its loop and describes the outcome qualitatively enters the lowest evidence tier (line 256).

more of objection 6

The implication is that a study with a failed loop is counted as missing certain evaluation designs. This should be made explicit in Section 5.1: a study with a negative result is included and its codes reflect what it reported, so the shares of reported designs are not inflated by only including successful studies.

What would settle it

Replicate the full-text coding on a sample of 50 self-improvement papers that do not use the RSI label (e.g., from the 'self-evolving agents' or 'iterative self-improvement' literature) to test whether the observed gap between product and process evaluation extends beyond the phrase-based corpus.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 21 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 31 s · 48,068 tokens read, 2,733 written
  3. Editor wrote the verdict above 11 s · 48,742 tokens read, 1,084 written
  4. Referee minor revision: 2 major objections, 6 minor 28 s · 48,315 tokens read, 2,412 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This is a solid piece of meta-research. The core contribution is straightforward but important: they actually count how often RSI papers report the evidence their claims require, and the gap between product evaluation (common) and process evaluation (rare) is convincing. The methodology is careful—reproducible arXiv search, full-text coding with locators, double coding with adjudication, an independent blind pass on the improver comparisons, and a comparison with Wu et al.'s anchor works. They also include formal propositions (null-loop benchmark, coupled audits, diagonal problem) that explain why product evidence cannot substitute for process evidence. The writing is direct and the limitations section is unusually honest.

The soft spots are real but mostly acknowledged. The phrase-based search frame captures only papers that use the RSI label, so the prevalence estimates may not generalize to all self-improving systems. The authors note this and check against anchor works, which are a selected set. The coding reliability is another concern: the first pass was AI-assisted and erred substantially—only the random sample (29 studies) received fully independent coding and adjudication. The rest use the fallible first-pass codes, though the robustness check shows no statistically significant differences on most items. The intervals are wide, so the null result is not strong. The independent improver coding covers only studies with a claim or a previously coded comparison, so missed comparisons could exist.

These caveats are in proportion: the main pattern (product evaluation >> process evaluation) holds in the adjudicated sample, the anchor works, and the rest of the frame. The paper does not overclaim; it presents counts with intervals and transparently discusses construct validity. The propositions are illustrative rather than definitive.

Who this is for: anyone working on or funding RSI research, meta-researchers studying evaluation practices, and journal editors who want a grounded standard for reporting. The paper deserves a serious referee and, with minor attention to framing the search limitation more prominently in the abstract, should be published. I would cite it.

The strongest objection

The paper's central finding—that process evaluation evidence is substantially less common than product evaluation in RSI studies—is robustly supported. The search-frame limitation is transparently acknowledged and mitigated by the anchor works and human-coded sample. The coding reliability is assessed via independent coding, adjudication, and separate blind coding of improver comparisons, with consistent results. The stricter definitions for process evidence (e.g., same-start comparisons, audited reruns) are justified by formal propositions. The paper does not overclaim; its central contribution is measuring reported practices, and it openly discusses construct validity concerns. No unhandled threat to the main conclusion is identified.

The reader's weakest_assumption—that the arXiv search frame may not generalize—is the most plausible concern, but the paper explicitly acknowledges this limitation (Section 9) and provides a robustness check via anchor works (20 representative studies outside the frame) that show the same pattern. The paper also presents human-coded random sample estimates that confirm the headline contrasts. The central claim that process evidence is rarer than product evidence is supported by multiple independent lines: (1) 22% audited reruns vs 90% broad product evaluation in the full set; (2) 14% vs 45% in the adjudicated sample; (3) similar patterns in anchor works (25% audited reruns vs 95% product evaluation). The improver comparison count (4 of 26) is based on independent blind coding with clarified decision rules, and the theoretical motivation (Proposition 3) justifies the strict same-start requirement. The paper's claims are measured and caveated appropriately. No additional load-bearing concern beyond those already discussed by the authors could be identified. Therefore the verdict should remain unchanged.

The referee's report

If the central finding holds—that process evaluation evidence in RSI research is substantially less common than product evaluation and often insufficient for the claims made—the paper provides an important empirical foundation for improving evaluation practice in a rapidly growing area. The methodology is careful: reproducible search, double coding with adjudication, independent blind coding of improver comparisons, comparison with anchor works, and explicit handling of coding reliability. The accompanying dataset and codebook are valuable resources for the community. The three propositions in Section 4 offer a formal basis for why product evidence cannot substitute for process evidence, though they are illustrative rather than central to the empirical results. The paper does not overclaim; its contribution is measurement and guidance, and it openly discusses construct validity and generalizability concerns.

Pith's reading of the paper (a machine's summary, not the authors')

Recursive self-improvement (RSI) promises AI systems that get better at producing further improvements. Evaluating that promise requires evidence about the process—does the loop repeat? does the improver actually improve? This paper audits the evidence in 90 arXiv papers that describe their own systems as RSI. The central finding is a gap: 90% of studies report product evaluation, but only 22% run two or more independent reruns of the full loop. Of the 26 studies that claim a better improvement procedure, only 4 compare an earlier and a later improver from identical starting artifacts at a matched budget. Three formal propositions explain why product evidence cannot substitute for process evidence. The paper derives a claim-indexed reporting standard and provides the full dataset of coded claims and evaluation designs.

What carries the argument. A codebook that maps five distinct claim types (product performance, accumulation, repeatability, improver gain, compounding) to observable evaluation designs, based on the evaluation principles of Wu et al. (2026) and Wang et al. (2026). The central objects are the distinction between development signal and audit, the swap test (comparing earlier and later improvers from identical starting artifacts), and the formal propositions that explain why product evidence cannot substitute for process evidence.

Core claim. The paper establishes a quantitative gap between the claims made by RSI studies and the evaluation designs they report. Using full-text coding of 114 eligible arXiv studies (90 self-described RSI systems), it finds that product evaluation is nearly universal (90% report an audit beyond development tasks under a broad criterion), but process evidence is far rarer. Only 22% audit two or more independent reruns of the loop, and only 4 out of 26 studies claiming an improver gain compare an earlier and a later improver from the same starting state at a matched budget. The gap persists in a human-coded random sample and in a comparison set of 20 established works. Three propositions—on selection bias, coupled audit errors, and the diagonal problem—show why product gains alone cannot establish repeatability or improver improvement.

The premise it rests on. The paper assumes that searching arXiv for papers whose title or abstract contains 'recursive self-improvement' (or a close variant) captures a meaningful population of self-improvement research; the authors acknowledge that work on similar mechanisms that does not use that phrase is omitted.

If the paper is right (Pith's inferences, not the authors')

If correct, the paper implies that most published RSI claims are supported only by product evidence, leaving the distinctively recursive claim of an improving improver untested. Readers should not infer a working RSI loop from a rising benchmark score alone.

The claim-indexed reporting standard (Table 9) gives authors a concrete checklist: state which claim you are making, and for each claim report the corresponding evaluation design. Widespread adoption would raise the bar for what counts as evidence for RSI.

The dataset and codebook enable future meta-researchers to measure whether evaluation practice improves over time, and to apply the same claim-level audit to related fields such as automated machine learning or self-evolving agents.

The formal propositions (null-loop benchmark, coupled audits, diagonal non-identification) give reviewers and authors a theoretical vocabulary for identifying when a reported evaluation design cannot support the claim it is paired with.

The gap this paper measures echoes earlier meta-research in machine learning: just as many deep RL papers reported point estimates from few runs, many RSI papers now report product gains without checking repeatability. The pattern suggests that computationally expensive loops lead researchers to skip reruns, even though reruns are essential for process claims.

The paper’s three propositions imply a concrete recommendation for RSI evaluation: a study that cannot afford repeated loop runs should restrict its claim to product performance and explicitly disclaim process claims like repeatability or improver gain. The codebook provides the vocabulary to make those restrictions explicit.

If the field adopts the swap test as standard, the cost of evaluation may rise, but the evidence for improver gains will become more trustworthy. Without such tests, the literature may accumulate apparent improver gains that reflect co-evolution of state and improver rather than a genuinely better improvement procedure.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Recursive self-improvement (RSI) promises AI systems, such as self-evolving LLM agents, that become better at producing further improvements. Evaluating this promise requires separating product claims about the final system from process claims about the improvement process itself. Existing surveys specify the evidence these claims require, but how often studies report that evidence remains unmeasured. We examine this question through full-text coding of 114 eligible studies in a reproducible arXiv search frame frozen in October 2026, including 90 that describe their own systems as RSI. The codebook links claims to evaluation designs. Independent human coding, third-reader adjudication, and a separate blind coding of improver comparisons assess coding reliability. Among the 90 RSI studies, 90% report product evaluation beyond development tasks under a broad criterion that includes partial or unstated separation, while 60% report evaluation on tasks unused for selection. Only 22% audit two or more reruns of the full loop. In a human-coded random sample, the RSI studies show the same contrast: 45% evaluate on tasks unused for selection and 14% audit two or more reruns. Of the 26 studies claiming a better improvement procedure, 23 report a related comparison, but only 4 compare earlier and later procedures from identical starting artifacts at a matched budget. Across these results and a comparison set of established works, evidence for repeatability and improved procedures remains less developed than product evaluation. Three propositions explain how selection effects, shared evaluator errors, and changes in system state limit what product evaluations can establish. Together, the findings support a claim-indexed reporting standard that connects each claim to the evaluation design needed to test it.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-21

The authors' main claim, as the classifier picked it out 0.56

“Across these results and a comparison set of established works, evidence for repeatability and improved procedures remains less developed than product evaluation.”

How the main claim is stated: asserted 0.93

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.29 · human participants 0.50 · controlled comparison 0.38 · randomized assignment 0.04 · states quantitative result 0.99 · compares with prior results 0.81 · mentions code or data availability 0.11

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Recursive self-improvement (RSI) promises AI systems, such as self-evolving LLM agents, that become better at producing further improvements. background 1.00 0.04 0.02 asserted 0.78
Evaluating this promise requires separating product claims about the final system from process claims about the improvement process itself. background 0.91 0.07 0.05 asserted 0.99
Existing surveys specify the evidence these claims require, but how often studies report that evidence remains unmeasured. gap 0.96 0.06 0.03 unclear 0.79
We examine this question through full-text coding of 114 eligible studies in a reproducible arXiv search frame frozen in October 2026, including 90 that describe their own systems as RSI. method 0.92 0.95 0.03 asserted 0.97
The codebook links claims to evaluation designs. method 0.97 0.78 0.02 asserted 0.98
Independent human coding, third-reader adjudication, and a separate blind coding of improver comparisons assess coding reliability. method 0.99 0.88 0.02 asserted 0.87
Among the 90 RSI studies, 90% report product evaluation beyond development tasks under a broad criterion that includes partial or unstated separation, while 60% report evaluation on tasks unused for selection. result 1.00 0.92 0.12 asserted 1.00
Only 22% audit two or more reruns of the full loop. result 1.00 0.87 0.08 asserted 1.00
In a human-coded random sample, the RSI studies show the same contrast: 45% evaluate on tasks unused for selection and 14% audit two or more reruns. result 1.00 0.91 0.07 asserted 1.00
Of the 26 studies claiming a better improvement procedure, 23 report a related comparison, but only 4 compare earlier and later procedures from identical starting artifacts at a matched budget. result 1.00 0.87 0.06 asserted 1.00
Across these results and a comparison set of established works, evidence for repeatability and improved procedures remains less developed than product evaluation. interpretation 0.51 0.82 0.46 asserted 0.93
Three propositions explain how selection effects, shared evaluator errors, and changes in system state limit what product evaluations can establish. interpretation 0.89 0.60 0.09 asserted 0.48
Together, the findings support a claim-indexed reporting standard that connects each claim to the evaluation design needed to test it. interpretation 0.58 0.51 0.75 asserted 0.73

What kind of paper the classifier reports this is: empirical study 0.58 · evidence from analysis of existing data 0.65

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-21: 1 request, 9,513 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (7)
Figure 1 of the paper
Figure 1. Reported evaluation designs in the 90 strong-label studies and 20 anchor works. The first two rows describe
Figure 3 of the paper
Figure 3. Lee et al. [76] Paired self-evaluation (same model) judge-graded gains; executable benchmarks flat (37.5 to 35.2, 62.6 to 61.1) Sec. 4;

Showing 2 of the paper's 7 figures. Figures are the authors', reproduced from the paper.

References (125)
  1. [1] Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents. InThe Fourteenth International Conference on Learning Representations, pages 104223–104294, 2026. URL https://proceedings.iclr.cc/paper_files/p... on Pith
  2. [2] Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems. InThe Thirteenth Interna- tional Conference on Learning Representations (ICLR), 2025. URL https://openreview.net/forum?id= t9U3LW7JVX on Pith
  3. [3] Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhuang, Yoonho Lee, Chengsong Huang, Han Yu, Zhongying CuiZhu, Yifei Ming, Huaxiu Yao, Burak Gokturk, Tomas Pfister, and Chen-Yu Lee. RRSI: regularized recursive Self-Improvement of agent harnesses. arXiv preprint arXiv:2609... on PithPith's reviewarXiv
  4. [5] Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason E Weston. Self-Rewarding language models. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pages 57... on Pith
  5. [6] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-Play Fine-Tuning converts weak language models to strong language models. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of Machine Learning Research, pag... on Pith
  6. [7] Shangjian Yin, Zehao Zhao, Kavosh Asadi, Rui Liu, Yuchen Lu, Shike Mei, Hang Cui, Luke Simon, Zhouxing Shi, and Hamed Firooz. Recursive Self-Improvement via On-Policy distillation for reasoning. arXiv preprint arXiv:2609.30652, 2026. URLhttps://arxiv.org/abs/2609.30652 on PithPith's reviewarXiv
  7. [8] Zihan Tan, Leixin Sun, Zitong Shi, Yitao Liu, Jiajun Wu, Nathaniel Brooks, Jiaru Qian, Xiaoran Shang, Suyuan Huang, Yi Ding, Yangxu Liao, Mukai Li, Qiushi Sun, Shudong Liu, Xuankun Rong, Xiaohang Yu, Zhuo Chen, Hejia Geng, Chenxin Li, Aozhou Wang, Zengji Tu, Robert Tang, Yuxin... on PithPith's reviewarXiv
  8. [9] Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng, Jiangwei Lao, Quanbin Wang, Honglin Lin, Jiajia Liu, Zhihong Lu, Hao Jin, Jian Wang, and Lijun Wu. ECG-RSI: recursive Self-Improvement through targeted electrocardiogram waveform evolution. arXiv preprint arXiv:2610.17718, 2026. URLh... on PithPith's reviewarXiv
  9. [10] Eric Zelikman, Eliana Lorch, Lester Mackey, and Adam Tauman Kalai. Self-Taught optimizer (STOP): recursively Self-Improving code generation. InFirst Conference on Language Modeling (COLM), 2024. URL https: //openreview.net/forum?id=46Zgqo4QIU on Pith
  10. [11] Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, and Tatiana Shavrina. Hyperagents. arXiv preprint arXiv:2603.19461, 2026. URL https://arxiv.org/abs/2603.19461 on PithPith's reviewarXiv
  11. [12] Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu, and Zhengyao Jiang. Recursive self-improvement of AI research agents. arXiv preprint arXiv:2609.26457, 2026. URLhttps://arxiv.org/abs/2609.26457 on PithPith's reviewarXiv
  12. [13] Tianhao Wu, Weizhe Yuan, Olga Golovneva, Jing Xu, Yuandong Tian, Jiantao Jiao, Jason E Weston, and Sainbayar Sukhbaatar. Meta-Rewarding language models: Self-Improving alignment with LLM-as-a-Meta- Judge. InProceedings of the 2025 Conference on Empirical Methods in Natural Lan... on Pithdoi:10.18653/v1/2025.emnlp-main.583
  13. [14] Alex Iacob, Andrej Jovanovi´c, William F. Shen, Daniel Burkhardt, Meghdad Kurmanji, Nurbek Tastan, Lorenzo Sani, Niccolò Alberto Elia Venanzi, Jiayi Nie, Ambroise Odonnat, Zeyu Cao, Bill Marino, Xinchi Qiu, Rika 21 Evaluation Evidence Behind RSI Claims Antonova, and Nicholas D... on PithPith's reviewarXiv
  14. [15] Chenyu Zhou. More convincing, not more correct: Self-Play reward hacking of Reference-Free LLM judges. arXiv preprint arXiv:2607.05904, 2026. URLhttps://arxiv.org/abs/2607.05904 on PithPith's reviewarXiv
  15. [16] Jiaming Wang. What stops recursive Self-Improvement in robotics? lessons from 123 rounds of agentic skill discovery. arXiv preprint arXiv:2609.31760, 2026. URLhttps://arxiv.org/abs/2609.31760 on PithPith's reviewarXiv
  16. [17] Hao Wu, Fan Xu, Fan Zhang, Zhipeng Xu, Yongheng Zhang, Yinghui Li, Yizhou Zhao, Penghao Zhao, Weiran Yao, Zengxiang Li, Yuan Gao, Zhangyang Gao, Hanbin Wang, Ruihan Tao, Wenjie Wang, Qi Li, Junfeng Fang, Xingjun Ma, Qiankun Li, Xiao Luo, Yefeng Zheng, Zhiwei Liu, Stephen Wang,... on Pithdoi:10.20944/preprints2026
  17. [18] Kaiqi Wang, Wenjin Hou, Yuchen Yan, Hongrui Jia, Zhisheng Zhong, Botao Ren, Yifei Chen, Songyang Zhang, Yongliang Shen, Jun Xiao, Yi Yang, Yueting Zhuang, and Hehe Fan. Diving into reliable Self-Evolving agents: A survey. Preprints.org, 2026. URLhttps://doi.org/10.20944/prepri... on Pithdoi:10.20944/preprints202609.0913.v1
  18. [19] Mingguang Chen, Licheng Wang, and Bo Qu. Recursive Self-Improvement in AI: from bounded Self-Refinement to autonomous research loops. arXiv preprint arXiv:2607.07663, 2026. URL https://arxiv.org/abs/2607. 07663 on PithPith's reviewarXiv
  19. [20] Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Qihan Ren, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, Cheng Qian, Zhenhailong Wan... on Pith
  20. [21] Jinyuan Fang, Yanwen Peng, Xi Zhang, Yingxu Wang, Xinhao Yi, Guibin Zhang, Yi Xu, Bin Wu, Siwei Liu, Zihao Li, Zhaochun Ren, Nikos Aletras, Xi Wang, Han Zhou, and Zaiqiao Meng. A comprehensive survey of Self-Evolving AI agents: A new paradigm bridging foundation models and lif... on PithPith's reviewarXiv
show 80 more
  1. [23] Zhe Ren, Yimeng Chen, Dandan Guo, Guowei Rong, Tonghui Li, R. B. Xiong, Qingfeng Lan, Wenyi Wang, Li Nanbo, Yibo Yang, Mingchen Zhuge, and Jürgen Schmidhuber. Self-Improvements in modern agentic systems: A survey. arXiv preprint arXiv:2607.13104, 2026. URLhttps://arxiv.org/abs... on PithPith's reviewarXiv
  2. [24] Leijun Zhou, Yanke Yu, Zhihao Liu, Qu Xiang, Tiancheng Hu, Yiru Dong, Yihan Wu, Yangxi Xiang, Xuejun Wu, Xi Chen, Buyue Qian, Yaowei Zheng, and Junhao Hu. From existing methods to new research ideas: Two perspectives on recursive Self-Improvement. arXiv preprint arXiv:2610.146... on PithPith's reviewarXiv
  3. [25] Zhipeng Liu, Xuefeng Bai, Kehai Chen, Xinyang Chen, Xiucheng Li, Yang Xiang, Jin Liu, Hong-Dong Li, Yaowei Wang, Liqiang Nie, and Min Zhang. A survey on the feedback mechanism of LLM-based AI agents. InThirty-Fourth International Joint Conference on Artificial Intelligence, pa... on Pithdoi:10.24963/ijcai.2025/1175
  4. [26] Jiahao Chen, Zhou Feng, Oubo Ma, Yichen Yan, Ruixiao Lin, Hangtao Zhang, Linkang Du, Hengyu An, Yong Yang, Jun Liu, Junhao Li, Naen Xu, Chunyi Zhou, Yuan Su, Zehao Jin, Qianli Ma, Leyi Qi, Yiming Wang, Zhe Ma, Yuwen Pu, Mengyao Du, Yuanyi Song, Enhao Huang, Zhihui Fu, Jun Wang... on PithPith's reviewarXiv
  5. [27] Chang Gong, Jingping Bi, Di Yao, Ran Cheng, Yaodong Yang, Kaile Wang, Xinjian Liang, Chao Xiang, and Ruijie Guo. Evolutionary safety of recursive Self-Improving AI: taxonomy, risk discovery, and evaluation. arXiv preprint arXiv:2609.31186, 2026. URLhttps://arxiv.org/abs/2609.31186 on PithPith's reviewarXiv
  6. [29] Hongyao Tang, Yi Ma, Pengyi Li, and Yifu Yuan. Generalized agent iteration: One formal framework for iterative policy improvement and recursive Self-Improvement. arXiv preprint arXiv:2609.13406, 2026. URL https://arxiv.org/abs/2609.13406 on PithPith's reviewarXiv
  7. [30] Chien-Ping Lu. Verification and Self-Improvement in agentic AI: foundations and limits. arXiv preprint arXiv:2610.10611, 2026. URLhttps://arxiv.org/abs/2610.10611 on PithPith's reviewarXiv
  8. [31] URLhttps://arxiv.org/abs/2609.11873 on PithPith's reviewarXiv
  9. [32] Kai Wu, Hao Lyu, Zhen Luo, Chaofan Wang, Siyu Ye, Jinghao Lin, Xiaozhong Ji, Boyuan Jiang, Shengzhi Wang, Zihan Wang, Yiwen Ye, Hao Wang, Zimu Wang, Wenzhe Liu, Ruobing Wang, Kai Cai, Mingliang Xiong, Wen Fang, Mingqing Liu, Yifan Zhang, Lei Yang, Xiaobin Hu, and Qingwen Liu. ... on Pithdoi:10.20944/preprints202608.2108.v1
  10. [33] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Kather- ine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-Refine: iterat... on Pith
  11. [34] Deheng Ye, Zheng Zhang, Hao Wang, and Chunyan Miao. A survey on AI for AI: when the improver becomes the improvee. Preprints.org, 2026. URLhttps://doi.org/10.20944/preprints202609.0591.v1 on Pithdoi:10.20944/preprints202609.0591.v1
  12. [35] Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. CRITIC: large language models can Self-Correct with Tool-Interactive critiquing. InThe Twelfth International Conference on Learning Representations (ICLR), 2024. URLhttps://openreview.net... on Pith
  13. [36] Ryo Kamoi, Yusen Zhang, Nan Zhang, Jiawei Han, and Rui Zhang. When can LLMs actually correct their own mistakes? a critical survey of Self-Correction of LLMs.Transactions of the Association for Computational Linguistics, 12:1417–1440, 2024. doi: 10.1162/tacl_a_00713. URL https... on Pithdoi:10.1162/tacl_a_00713
  14. [37] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automat- ically correcting large language models: Surveying the landscape of diverse automated correction strategies. Transactions of the Association for Computational Linguistics, 12:48... on Pithdoi:10.1162/tacl_a_00660
  15. [38] Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot Self-Correct reasoning yet. InThe Twelfth International Conference on Learning Representations (ICLR), 2024. URLhttps://openreview.net/forum?i... on Pith
  16. [39] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan P... on PithPith's reviewarXiv
  17. [40] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: language agents with verbal reinforcement learning. InAdvances in Neural Information Processing Systems, volume 36, pages 8634–8652, 2023. doi: 10.52202/075280-0377. URL https://paper... on Pithdoi:10.52202/075280-0377
  18. [41] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anand- kumar. V oyager: An Open-Ended Embodied Agent with Large Language Models.Transactions on Machine Learning Research, 2024. URLhttps://mlanthology.org/tmlr/2024/wang2024tmlr... on Pith
  19. [42] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. STaR: bootstrapping reasoning with reasoning. InAdvances in Neural Information Processing Systems, volume 35, pages 15476–15488, on Pith
  20. [43] Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. AI models collapse when trained on recursively generated data.Nature, 631(8022):755–759, 2024. doi: 10.1038/s41586-024-07566-y on Pithdoi:10.1038/s41586-024-07566-y
  21. [44] Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G. Baraniuk. Self-Consuming generative models go MAD. InThe Twelfth International Conference on Learning Representations (ICLR), 2024. URL ht... on Pith
  22. [45] Seungyoo Lee, Giung Nam, Hyungi Lee, and Juho Lee. OCNR: stabilizing Self-Play by mitigating Iteration- Collapse with One-Class novelty rewards. InProceedings of the 43rd International Conference on Machine Learning, volume 306 ofProceedings of Machine Learning Research, pages... on Pith
  23. [46] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph Gonzalez, and Ion Stoica. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. InAdvances in Neural Information Processing Sys... on Pithdoi:10.52202/075280-2020
  24. [47] Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. ExpeL: LLM Agents Are Experiential Learners. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19632–19642, 2024. doi: 10.1609/aaai.v38i17.29936. URL https://ojs.a... on Pithdoi:10.1609/aaai.v38i17.29936
  25. [48] Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, and Hannaneh Hajishirzi. RewardBench: evaluating reward models for language modeling. InFindings of the Association f... on Pithdoi:10.18653/v1/2025.findings-naacl.96
  26. [49] Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. InProceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of Machine Learning Re- search, pages 10835–10866. PMLR, 2023. URLhttps://proceedings.mlr.... on Pith
  27. [50] Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and characterizing reward gaming. InAdvances in Neural Information Processing Systems, volume 35, pages 9460–9471, on Pith
  28. [51] Carson Denison, Monte MacDiarmid, Fazl Barez, David Duvenaud, Shauna Kravec, Samuel Marks, Nicholas Schiefer, Ryan Soklaski, Alex Tamkin, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, Ethan Perez, and Evan Hubinger. Sycophancy to subterfuge: Investigating Reward-Tampering in... on PithPith's reviewarXiv
  29. [52] Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Lingpeng Kong, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1... on Pithdoi:10.18653/v1/2024.acl-long.511
  30. [53] Cynthia Dwork, Vitaly Feldman, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Aaron Roth. The reusable holdout: Preserving validity in adaptive data analysis.Science, 349(6248):636–638, 2015. doi: 10.1126/science. aaa9375 on Pithdoi:10.1126/science
  31. [54] Avrim Blum and Moritz Hardt. The ladder: A reliable leaderboard for machine learning competitions. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 ofProceedings of Machine Learning Research, pages 1006–1014. PMLR, 2015. URL https://proceeding... on Pith
  32. [55] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? InProceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings 24 Evaluation Evidence Behind RSI Claims of Machine Learning R... on Pith
  33. [56] doi: 10.52202/068431-0687. URL https://papers.nips.cc/paper_files/paper/2022/hash/ 3d719fee332caa23d5038b8a90e81796-Abstract-Conference.html on Pithdoi:10.52202/068431-0687
  34. [57] Xiaojing Sun, Yuhan Zeng, Zihua She, and Xiao Wang. Which Self-Improvements should we trust? reliable Self-Improvement when agents reuse their benchmarks. arXiv preprint arXiv:2609.33180, 2026. URL https: //arxiv.org/abs/2609.33180 on PithPith's reviewarXiv
  35. [58] Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, Ilya Sutskever, and Jeffrey Wu. Weak-to-Strong generalization: Eliciting strong capabilities with weak supervision. InProcee... on Pith
  36. [59] Fanqing Meng, Lingxiao Du, Qiguang Chen, Ziqi Zhao, Haocheng Lu, Mengkang Hu, and Michael Qizhe Shieh. RSIBench-Data: benchmarking Data-Centric research for recursive Self-Improvement. arXiv preprint arXiv:2607.25886, 2026. URLhttps://arxiv.org/abs/2607.25886 on PithPith's reviewarXiv
  37. [60] Yizhe Chi, Wenyi Li, Deyao Hong, Xiaoqiu Wang, Mingju Gao, Kaisen Yang, Bingxiang He, Youjie Zheng, Calvin Xiao, and Qinhuai Na. AI4AI-Bench: benchmarking LLM agents in algorithmic design for recursive Self-Improvement. arXiv preprint arXiv:2608.20318, 2026. URLhttps://arxiv.o... on PithPith's reviewarXiv
  38. [61] Jingwei Song, Zhisheng Chen, Bingfan Zeng, Jing Tang, Yubo Hou, Jinhao Jing, Zhengwei Xie, and Linfeng Zhang. MC-EvoBench: auditing Experience-Driven Self-Improvement in Minecraft agents. arXiv preprint arXiv:2610.13829, 2026. URLhttps://arxiv.org/abs/2610.13829 on PithPith's reviewarXiv
  39. [62] Rebecca Roelofs, Vaishaal Shankar, Benjamin Recht, Sara Fridovich-Keil, Moritz Hardt, John Miller, and Ludwig Schmidt. A Meta-Analysis of overfitting in machine learning. InAdvances in Neural Information Processing Systems, volume 32, 2019. URL https://papers.nips.cc/paper_fil... on Pith
  40. [63] Renxiong Wang, Darvin Yi, Abril Herrlein, Anas Mahmoud, Advait Gosai, Lisiman Hua, MohammadHossein Rezaei, Xingang Guo, Anisha Gunjal, Utkarsh Tyagi, David J. Lee, Minglai Yang, Haris Riaz, Chenguang Wang, Huaxiu Yao, Daniel Yue Zhang, Aakash Sabharwal, Tong Zhao, and Yunzhong... on PithPith's reviewarXiv
  41. [64] Shuhan Xue, Zixin Ding, Yichen Shen, Yinjie Wang, Zhenfei Yin, Yingcheng Wu, Yuxin Chen, Mengdi Wang, and Ling Yang. PAST-Bench: benchmarking the foundations of recursive Self-Improvement in personal agents. arXiv preprint arXiv:2608.04003, 2026. URLhttps://arxiv.org/abs/2608.04003 on PithPith's reviewarXiv
  42. [66] Odd Erik Gundersen and Sigbjørn Kjensmo. State of the art: Reproducibility in artificial intelligence. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. doi: 10.1609/aaai.v32i1.11503. URLhttps://ojs.aaai.org/index.php/AAAI/article/view/11503 on Pithdoi:10.1609/aaai.v32i1.11503
  43. [67] Benjamin Marie, Atsushi Fujita, and Raphael Rubino. Scientific credibility of machine translation research: A Meta-Evaluation of 769 papers. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on... on Pithdoi:10.18653/v1/2021.acl-long.566
  44. [68] Fanqing Meng, Lingxiao Du, Haocheng Lu, Qiguang Chen, Ziqi Zhao, Zijian Wu, Jiayuan Zhuo, Mengkang Hu, and Michael Qizhe Shieh. RSIGym: a flexible environment for recursive Self-Improvement. arXiv preprint arXiv:2610.10310, 2026. URLhttps://arxiv.org/abs/2610.10310 on PithPith's reviewarXiv
  45. [69] Rishabh Agarwal, Max Schwarzer, Pablo Samuel Castro, Aaron Courville, and Marc Bellemare. Deep reinforce- ment learning at the edge of the statistical precipice. InAdvances in Neural Information Processing Systems, volume 34, pages 29304–29320, 2021. URL https://papers.nips.cc... on Pith
  46. [70] Yuxuan Liu, Zhaochen Su, Yuhao Zhang, Jiahe Guo, Zhongwei Xie, Huihao Jing, Lingyun Xie, Qing Zong, Yauwai Yim, Zhixiong Zhang, Haoran Li, and Yangqiu Song. Rethinking Self-Evolving agent skills: Feedback dynamics over multiple rounds. arXiv preprint arXiv:2608.02636, 2026. UR... on PithPith's reviewarXiv
  47. [71] Sayash Kapoor and Arvind Narayanan. Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9):100804, 2023. doi: 10.1016/j.patter.2023.100804 on Pithdoi:10.1016/j.patter.2023.100804
  48. [72] URLhttps://arxiv.org/abs/2609.35596 on PithPith's reviewarXiv
  49. [73] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-Instruct: aligning language models with Self-Generated instructions. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics ... on Pith
  50. [74] Qianru Meng, Xi Zhang, Konstantinos Papakostas, Congfeng Cao, Wenxiang Hu, Tianyi Lu, Xiao Zhang, Lun Zhou, Zaiqiao Meng, Zhaochun Ren, Xiaohui Yan, and Yougang Lyu. SwarmRSI: recursive Self-Improvement through evolving scientific agent swarm. arXiv preprint arXiv:2610.20431, ... on PithPith's reviewarXiv
  51. [75] Peter Henderson, Riashat Islam, Philip Bachman, Joelle Pineau, Doina Precup, and David Meger. Deep reinforcement learning that matters. InProceedings of the AAAI Conference on Artificial Intelligence, volume 32, on Pith
  52. [76] Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, Matei Zaharia, and Yujin Tang. Recursive Self-Improvement through Multi-Agent Self-Supervision. arXiv preprint arXiv:2610.12176, 2026. URLhttps://arxiv.org/abs/2610.12176 on PithPith's reviewarXiv
  53. [77] Yoichi Ishibashi, Taro Yano, and Masafumi Oyamada. Can Large Language Models Invent Algorithms to Improve Themselves?: Algorithm Discovery for Recursive Self-Improvement through Reinforcement Learning. In Proceedings of the 2025 Conference of the Nations of the Americas Chapte... on Pith
  54. [78] Jesse Dodge, Suchin Gururangan, Dallas Card, Roy Schwartz, and Noah A. Smith. Show your work: Improved reporting of experimental results. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu... on Pithdoi:10.18653/v1/d19-1224
  55. [79] Huan Zhang, Mingju Chen, Dongxu Zhou, Can Lv, Sen Cui, Faguo Wu, and Shiji Zhou. RSIFlow: adaptive Self-Improvement workflows with an evolving Meta-Agent. arXiv preprint arXiv:2610.21791, 2026. URL https://arxiv.org/abs/2610.21791 on PithPith's reviewarXiv
  56. [80] Wenyi Wang, Piotr Pi˛ ekos, Li Nanbo, Firas Laakom, Yimeng Chen, Mateusz Ostaszewski, Mingchen Zhuge, and Jürgen Schmidhuber. Huxley-Gödel Machine: Human-Level coding agent development by an approximation of the optimal Self-Improving machine. arXiv preprint arXiv:2510.21614, ... on PitharXiv
  57. [81] Haoyu Dong, Yuhang Zhou, Zihao Lin, Yifan Wu, Bo Peng, Mingyi Wang, Xiangjun Fan, Lizhu Zhang, and Zhuokai Zhao. Mixture of Self-Improving branches for agent harness optimization. arXiv preprint arXiv:2609.37834, 2026. URLhttps://arxiv.org/abs/2609.37834 on PithPith's reviewarXiv
  58. [82] doi: 10.18653/v1/2023.acl-long.754. URLhttps://aclanthology.org/2023.acl-long.754/ on Pithdoi:10.18653/v1/2023.acl-long.754
  59. [84] Feng Yuan, Xinyi Gao, Zijun Sun, Haoyue Li, Menglei Hua, Xiao Li, and Yifan Gao. Sivia: Learning to draw, learning to learn. arXiv preprint arXiv:2609.33386, 2026. URLhttps://arxiv.org/abs/2609.33386 on PithPith's reviewarXiv
  60. [85] Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa, and Dongyeop Kang. Metan: Recursive Self-Improvement through emergent depth. arXiv preprint arXiv:2608.24735, 2026. URL https://arxiv.org/abs/2608. 24735 on PithPith's reviewarXiv
  61. [86] Zipei Yu, Yue-Jiao Gong, Zeyuan Ma, Yuncheng Jiang, and Zhiguang Cao. Hyper algorithm design agent: Evolving learnable optimizer from zero. arXiv preprint arXiv:2609.35328, 2026. URLhttps://arxiv.org/ abs/2609.35328. 26 Evaluation Evidence Behind RSI Claims on PithPith's reviewarXiv
  62. [87] Xin Xu, Ke Cheng, Hao Wang, Yangru Huang, Yanghao Xiao, Zhichao Chen, Zhisheng Zhong, Xun Liu, Yi Li, Qian Li, Haibin Lv, Jun Zhang, Jie Jiang, and Zhouchen Lin. Drafting agents, not just tokens: Speculative recursive Self-Improvement for agentic systems. arXiv preprint arXiv:... on PithPith's reviewarXiv
  63. [88] Xunjian Yin, Xinyi Wang, Liangming Pan, Li Lin, Xiaojun Wan, and William Yang Wang. Gödel Agent: A Self-Referential Agent Framework for Recursively Self-Improvement. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Paper... on Pithdoi:10.18653/v1/2025.acl-long.1354
  64. [89] Tong Zheng, Xidong Wu, Zheng Zhang, Zhankui He, Chaoyi Zhang, Benjamin Coleman, Ruoqiao Wei, Di Bai, Haolin Liu, Rui Liu, Xue Wang, Yue Zhuan, Wang-Cheng Kang, Renkai Xiang, Heng Huang, Xinwu Cheng, and Yunsong Guo. Dream-RSI: recursive Self-Improvement through evolving worlds... on PithPith's reviewarXiv
  65. [90] Kaiyuan Liu, Ziyuan Zhuang, Rongxiang Weng, and Jieping Ye. RecurSE: bounded recursive Self-Evaluation for LLM rubric judges. arXiv preprint arXiv:2608.24231, 2026. URLhttps://arxiv.org/abs/2608.24231 on PithPith's reviewarXiv
  66. [91] Wei Dai, Qianchu Liu, Guanghui Qin, Xing Han, Tristan Naumann, Paul Pu Liang, and Sheng Zhang. Recursive reward improvement via adversarial Meta-Prompt evolution. arXiv preprint arXiv:2610.21040, 2026. URL https://arxiv.org/abs/2610.21040 on PithPith's reviewarXiv
  67. [92] Jiacheng Guo, Suozhi Huang, Yunlong Gao, Zihao Li, Jason Ge, Shushu Liang, Xu Kuang, and Mengdi Wang. AQuA: recursively Self-Improving quantitative trading research agents. arXiv preprint arXiv:2608.12841, 2026. URLhttps://arxiv.org/abs/2608.12841 on PithPith's reviewarXiv
  68. [93] URLhttps://arxiv.org/abs/2607.05297 on PithPith's reviewarXiv
  69. [94] Wenqian Ye, Ziwei Guan, Zi Yang, Shivani Modi, Cici Cong Wu, Yunlong He, Yuxin Chen, Zhengyu Zhang, Liang Luo, Venkatesh Ranganathan, Jianhui Sun, Yantao Yao, Wenlin Chen, Buyun Zhang, Ellie Dingqiao Wen, Henry Kautz, and Aidong Zhang. Heterogeneous agentic coevolution. arXiv ... on PithPith's reviewarXiv
  70. [96] Architect Labs. Redwood: A frontier AI accelerator designed, verified, and deployed from scratch in 2 weeks by AI. arXiv preprint arXiv:2608.26418, 2026. URLhttps://arxiv.org/abs/2608.26418 on PithPith's reviewarXiv
  71. [97] Taoyong Cui, Pinxian Zeng, Wanghan Xu, Hao Xiao, Shuhan Xue, Qiang Gao, Wanli Ouyang, Pheng Ann Heng, Zhenfei Yin, Yingcheng Wu, and Ling Yang. RSI-Anything: unlocking recursive Self-Improvement with world models. arXiv preprint arXiv:2610.13912, 2026. URLhttps://arxiv.org/abs... on PithPith's reviewarXiv
  72. [98] Changzhi Liu, Yilun Liu, Sikuan Yan, V olker Tresp, and Yunpu Ma. Mendel Gödel Machine: Recursive Self-Improving coding agents via comparative evolution. arXiv preprint arXiv:2608.07645, 2026. URL https://arxiv.org/abs/2608.07645 on PithPith's reviewarXiv
  73. [99] Aditya Namdev Kakade, Vivek Srivastava, and Shirish Karande. POLARIS: A Gödel Agent Framework for Small Language Models through Experience-Abstracted Policy Repair. InFindings of the Association for Computational Linguistics: ACL 2026, pages 39505–39537, 2026. doi: 10.18653/v1... on Pithdoi:10.18653/v1/2026.findings-acl.1969
  74. [100] Zihao Lin, Haoyu Dong, Haobo Xu, Chi-Chih Chang, Xuying Ning, Yinglong Xia, Shuang Yang, Zhuokai Zhao, Xinying Song, Dongfang Liu, Lifu Huang, Yang Song, and Qifan Wang. VideoHarness-SRSI: structured recursive Self-Improvement for video understanding harness. arXiv preprint ar... on PithPith's reviewarXiv
  75. [102] URLhttps://arxiv.org/abs/2610.08640 on PithPith's reviewarXiv
  76. [103] Yuxiang Wang, Shengbo Cai, Yingda Shen, Ming-Hao Hsu, Qinke Ni, Liqiang Zhang, Teddy Sun, Steve Yevs, and Zhizheng Wu. EvoAudio: recursive Self-Improvement for audio understanding. arXiv preprint arXiv:2609.27389, 2026. URLhttps://arxiv.org/abs/2609.27389 on PithPith's reviewarXiv
  77. [104] Shuxing Yang, Kaihao Zhu, Junjie Yang, Rui Zhao, Junyao Wu, Yize Wang, Wenhao Li, Fujia Chen, Taowen Deng, Shenzhan Hong, Yaqi Li, Zichen Li, Jincheng Mi, Yuang Pan, Hongsheng Chen, and Yihao Yang. Data- Efficient language modeling: From frontier advancement to Principle-Guide... on PithPith's reviewarXiv
  78. [105] Wenqing Tian, Zeyu Zhang, Zhaocheng Liu, Fengwei Liu, Qiang Liu, and Liang Wang. PhysEvo: astra can act, let it. arXiv preprint arXiv:2610.08995, 2026. URLhttps://arxiv.org/abs/2610.08995 on PithPith's reviewarXiv
  79. [106] Zhihui Chen and Mengling Feng. MedForge-RSI: medical deepfake detection via recursive Self-Improvement. arXiv preprint arXiv:2609.36549, 2026. URLhttps://arxiv.org/abs/2609.36549 on PithPith's reviewarXiv
  80. [107] URLhttps://arxiv.org/abs/2607.28568 on PithPith's reviewarXiv

The first 100 of 125 references are listed here. All of them are in this paper's citation record.

Cited by 1 Pith paper

Reviewed papers on Pith that cite this one. The full citation record

  1. Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents cs.AI · Nov 2026

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment