Artificial Intelligence · arXiv:2611.23138 v1 · submitted 17 Nov 2026 · 52 references
RSI-Prime: Training Models to Train Models
The paper says
“Enabling large language models (LLMs) to autonomously conduct training and improve target models is an important foundation for recursive self-improvement (RSI).”
more of the abstract
As LLMs expand their role from being trained to conducting training, their model-training capability warrants dedicated study and systematic development. To this end, we construct RSI-Prime, a dataset for improving model-training capability, comprising 350 complete experimental trajectories across seven task categories. To help LLMs learn to make high-quality experimental decisions from training practice, we perform fine-grained decision identification and quality annotation, then apply decision-level loss masking to exclude supervision on low-quality behavior. Experiments with GLM-5.3 on PostTrainBench, a benchmark for autonomous post-training, show that training on RSI-Prime yields a 31.95% relative improvement in overall performance over the original GLM-5.3, highlighting a path toward future RSI systems in which models autonomously sustain their own advancement.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 18 Nov 2026 · read the full text (70,127 characters) · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A well-executed study on using decision-level supervision to improve LLMs' model-training capability, with a clear result on PostTrainBench but some open questions about evaluation overlap and reproducibility. This line is wrong
10 minor · 1 open question
How often objections from reviews like this one are right: not yet measured
The paper, first character to last
The reviewer was given the whole text, 70,127 characters. Not checked: figures, code, what the 52 references say.
10 minor objections
The paper says, at §1, Abstract
Pith says: minor objection 1 of 10
The abstract in the manuscript matches the arXiv metadata abstract, which is good. However, the abstract states '350 complete experimental trajectories across seven task categories' while Appendix A.1 says 'The dataset comprises 350 benchmarks, 350 trajectories, and 385 training segments.' The relationship between benchmarks, trajectories, and training segments is unclear. Clarify whether each benchmark corresponds to one trajectory or if some benchmarks have multiple trajectories.
The paper says, at §3.2, §3.3
Pith says: minor objection 2 of 10
The paper mentions 'seven task categories' but does not list them explicitly in the main text. They are only implied through the benchmark selection description. A clear enumeration of the seven categories would improve readability.
show 8 more
The paper says, at §4, Table 2
Pith says: minor objection 3 of 10
The table caption says 'Superscripts report relative gains in Overall over the original GLM-5.3 agent.' However, the superscripts show '+24.75%' and '+31.95%' which are relative gains, but the table does not show the original GLM-5.3 Overall score in the same row for easy comparison. Consider adding a column or note showing the baseline score.
The paper says, at §4.3, Figure 3 (Figure~\ref{fig:agent-behavior})
Pith says: minor objection 4 of 10
The figure caption mentions '20 secondary action types' but the figure itself is not visible in the manuscript. The text references specific action types (e.g., 'learning objective, data source selection, and inference configuration') but does not provide a complete list. Consider adding a table with all 20 action types and their pass rates for each supervision strategy.
The paper says, at §3.6, Appendix A.3
Pith says: minor objection 5 of 10
The paper states 'Masked token counts therefore do not estimate the amount of erroneous content' (Appendix A.3). This is an important caveat that should be mentioned earlier, perhaps in §3.6 where the masking is introduced, to avoid misinterpretation of the retention statistics.
The paper says, at §4, Table 2
Pith says: minor objection 6 of 10
The table shows results for 'GLM-5.3 & DSH' with Overall 36.56, but Appendix B (Table 5) reports the mean HumanEval score as 79.27, while the main table shows 79.27 for GLM-5.3. However, the main table also shows 'GLM-5.2 & Claude Code' with HumanEval 74.66. It would be helpful to explain why GLM-5.3 with DSH is chosen as the primary baseline rather than GLM-5.2 with Claude Code which has a higher Overall score (39.37 vs 36.56).
The paper says, at §3.5
Pith says: minor objection 7 of 10
The paper uses 'GLM-5.3' for quality assessment but does not specify whether this is the same checkpoint as the one being fine-tuned or a different version. Clarify if the quality assessment uses the original GLM-5.3 or the fine-tuned version.
The paper says, at §2.2
Pith says: minor objection 8 of 10
The related work section mentions RSI-Dojo (Gao et al., 2026b) which also constructs post-training environments and collects trajectories. A more detailed comparison of RSI-Prime with RSI-Dojo in terms of dataset size, annotation granularity, and results would help position the contribution.
The paper says, at §3.4, Appendix A.4
Pith says: minor objection 9 of 10
The human audit (Appendix A.4) samples 400 decision records but 'does not estimate the recall of decision identification or the frequency of omitted decisions.' This is a significant limitation for the decision identification step, as missed decisions could affect the quality of supervision selection. The paper should discuss this limitation more prominently.
The paper says, at §4.1
Pith says: minor objection 10 of 10
The paper states 'Decision SFT improves over the original GLM-5.3 agent on all six tasks.' However, the original GLM-5.3 scores in Table 2 show HumanEval at 79.27, while Decision SFT achieves 83.54. Given the acknowledged training-evaluation overlap for HumanEval, this claim should be qualified.
Open questions (1)
Pith raised these and could not confirm them against the text of the paper it read. They are questions for the authors, not findings.
- §4.2, Table 3 (Table~\ref{tab:supervision-strategies}) Pith raised a concern here that it could not check against the text of the paper it was given, so it is not published as an objection.
What would settle it
If an independent replication using the same RSI-Prime dataset and training recipe found that decision-level supervision does not consistently outperform full-trajectory fine-tuning across multiple random seeds or base models, or that the performance gains are primarily driven by the HumanEval content overlap rather than the decision-level supervision mechanism, the central claim would be weakened.
from Pith's summary of the paper
How the review went
5 passes by deepseek-v4-flash · 18 Nov 2026
- Reader holds up, with conditions
- Skeptic concern withdrawn: Pith's reviewer misread the paper
- Referee 10 minor
- Check against the paper kept 0 of 5 objections
- Editor wrote the verdict above
Not run on this paper: the simulated authors' defence.
Pith's passes raised 5 objections. A check against the paper's text kept 0, withdrew 4 (1 the paper already says; 3 Pith's reviewer misread the paper) and left 1 as a question.
The review in full
The review in brief, machine-written
This paper introduces RSI-Prime, a dataset of 350 complete model-training trajectories with fine-grained decision-level annotations, and shows that training GLM-5.3 on this dataset with decision-level loss masking yields a 31.95% relative improvement on PostTrainBench v1.2. The core idea—identifying and selectively supervising high-quality experimental decisions within trajectories—is sensible and the empirical results are reasonably convincing. The paper is clearly written and the methodology is well described.
The main strengths are the careful construction of the dataset, the systematic pipeline for decision identification and quality annotation, and the ablation studies comparing different supervision strategies. The human audit of annotations (97.5-99.5% acceptance) adds credibility. The behavioral analysis showing improved decision quality across action types is a nice addition.
The soft spots are real but not fatal. The training corpus contains HumanEval-related content, creating a training-evaluation overlap that the paper acknowledges but does not fully quantify. The main results lack error bars—only the baseline GLM-5.3 was run multiple times, not the fine-tuned models. The quality assessment uses GLM-5.3 itself (with majority voting and human audit), which introduces some circularity. The dataset and code are not released, limiting reproducibility. These are common issues in this area but they temper confidence in the exact magnitude of the reported gain.
This paper is for researchers working on autonomous model training, self-improving agents, and trajectory learning. It deserves a serious peer review. The central claim is plausible and the methodology is sound enough to warrant further scrutiny. I would recommend engaging with the work while pressing for error bars, clarification of the HumanEval overlap's impact, and ideally a commitment to release the dataset.
The stress test's concern
Withdrawn: Pith's reviewer misread the paper.
The paper says
“This three-run averaging applies only to the original GLM-5.3 agent.”
The referee's report
Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 4 of its 4 major objections did not survive that check.
Pith's reading of the paper (a machine's summary, not the authors')
The paper aims to establish that large language models can improve their own ability to train other models by learning from high-quality experimental decisions extracted from complete training trajectories. The authors construct RSI-Prime, a dataset of 350 complete model-training experiments across seven task categories, with fine-grained identification and quality annotation of each experimental decision. They apply decision-level loss masking to supervise only high-quality behavior while preserving the full experimental context. Fine-tuning GLM-5.3 on this dataset yields a 31.95% relative improvement on PostTrainBench v1.2 (score 48.24 vs 36.56), and behavioral analysis shows improvements in decision quality across multiple stages of experimentation. If this approach generalizes, it could provide a path toward recursive self-improvement where models autonomously sustain their own advancement by learning from their own training practice.
What carries the argument. The central mechanism is decision-level loss masking applied to supervised fine-tuning. The pipeline first identifies experimental decisions within complete trajectories and annotates each decision along three quality dimensions (Analysis, Choice, Execution) using GLM-5.3 with majority voting and human audit. Decisions are mapped to assistant messages via source references; a message is supervised only if all its linked decisions pass quality selection (all three dimensions labeled reasonable or uncertain). The resulting mask is applied to the standard next-token prediction loss, preserving all original messages and tool interactions but excluding tokens from messages associated with low-quality decisions. This allows the model to learn from high-quality behavior while still observing the full experimental context including failures and corrections.
Core claim. The paper claims that training an LLM agent (GLM-5.3) on a curated dataset of complete model-training trajectories, with selective supervision guided by fine-grained decision quality annotations, substantially improves the agent's ability to autonomously conduct post-training of target models. The central discovery is that decision-level loss masking—excluding supervision on low-quality experimental decisions while preserving the full trajectory context—outperforms both full-trajectory fine-tuning and alternative selection strategies (tool-failure masking, step-level masking) at similar supervision volumes. On PostTrainBench v1.2, decision-level supervision achieves an overall score of 48.24, a 31.95% relative gain over the original GLM-5.3 agent (36.56), and surpasses full-trajectory fine-tuning (45.61) on all six tasks. Behavioral analysis further shows that decision-level supervision improves the quality of experimental decisions across 18 of 20 action types, with particularly large gains in key training choices such as learning objective selection and data source selection.
The premise it rests on. The decision quality annotations, generated by GLM-5.3 with majority voting and a human audit on 400 records, reliably distinguish high-quality from problematic experimental decisions, and the mapping from dimension-level judgments to binary quality labels captures meaningful distinctions that generalize across tasks and agent behaviors.
If the paper is right (Pith's inferences, not the authors')
If the approach scales, LLM agents could improve their model-training capability by learning from their own past experiments, reducing the need for human demonstration data or hand-crafted training curricula.
The decision-level supervision framework could be applied to other long-horizon agent tasks beyond model training, such as scientific research, software engineering, or robotics, where final success does not guarantee the quality of all intermediate decisions.
The 31.95% relative improvement on PostTrainBench suggests that current agents have substantial room to improve their experimental decision-making, and that learning from training experience is a viable path toward closing the gap with top-performing closed-source agents.
The finding that decision-level supervision outperforms step-level masking at similar supervision volumes implies that the granularity and criterion of supervision selection matter more than the amount of supervision retained.
The paper's approach implicitly assumes that the quality annotations capture meaningful distinctions that generalize across tasks and agent behaviors. A natural extension would be to test whether the same annotation pipeline and training recipe transfer to different base models (e.g., GLM-5.2, Kimi K3) or to different agent harnesses (e.g., Claude Code, Codex CLI) without retraining the annotation
The decision-level loss masking scheme is conservative: it masks entire messages if any linked decision is low-quality, which may discard useful contextual information. An alternative approach could apply token-level or span-level masking within messages, potentially retaining more high-quality content while still excluding problematic behavior.
The paper's trajectory filtering criterion (Eq. 1) requires the submitted model's improvement to exceed half the reference Instruct–Base gap. This threshold is somewhat arbitrary; varying it could affect the diversity and difficulty of trajectories in the dataset, and the sensitivity of the method to this threshold is not explored.
The human audit samples 400 decision records but does not estimate recall of decision identification or frequency of omitted decisions. A more thorough evaluation of the annotation pipeline's completeness would strengthen confidence that the dataset covers all relevant experimental decisions.
Check this review
What was read
The passes were given all 70,127 characters of the paper's LaTeX source.
Fingerprint of that text (SHA-256): db7157d3cee6696c12ca6ffe1606e7348cc5259759394c54655279238092820d
Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:
PYTHONPATH=. python scripts/review_input_check.py 2611.23138
The passes, in the order they ran
- Reader deepseek-v4-flash · medium effort · $0.0028 · instructions of the Reader, fingerprint f9ead7300382…
- Skeptic deepseek-v4-flash · medium effort · $0.0021 · instructions of the Skeptic, fingerprint 8f88c0f61488…
- Referee deepseek-v4-flash · medium effort · $0.0025 · instructions of the Referee, fingerprint a852d99aea0d…
- Check against the paper deepseek-v4-flash · medium effort · $0.0026 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
- Editor deepseek-v4-flash · medium effort · $0.0022 · instructions of the Editor, fingerprint e7211f6a3d27…
- Circularity check deepseek-v4-flash · medium effort · $0.0021 · instructions of the Circularity check, fingerprint e9946e39d67b…
- Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0025 · instructions of the Summary for the page, fingerprint e4f56bf568e7…
The passes cost $0.017 in model calls together.
Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00
Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.
curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum
prints the fingerprint of the reader's instructions. The same works for every pass.
The receipt
The review's fingerprint was taken on 18 Nov 2026. The review you are reading still matches it.
Fingerprint of the review (SHA-256): 56bfd58b554aac42327cbf2fc16c648dd4af7a914a03edbc3ff7e913aaa8aa1a
curl -s 'https://pith.science/api/v1/verdicts/2611.23138?form=canonical' | sha256sum
prints the fingerprint of the review as it stands now. The rule, in four steps.
A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.
How often objections from reviews like this one are right: not yet measured
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.
The paper
Abstract
Enabling large language models (LLMs) to autonomously conduct training and improve target models is an important foundation for recursive self-improvement (RSI). As LLMs expand their role from being trained to conducting training, their model-training capability warrants dedicated study and systematic development. To this end, we construct RSI-Prime, a dataset for improving model-training capability, comprising 350 complete experimental trajectories across seven task categories. To help LLMs learn to make high-quality experimental decisions from training practice, we perform fine-grained decision identification and quality annotation, then apply decision-level loss masking to exclude supervision on low-quality behavior. Experiments with GLM-5.3 on PostTrainBench, a benchmark for autonomous post-training, show that training on RSI-Prime yields a 31.95% relative improvement in overall performance over the original GLM-5.3, highlighting a path toward future RSI systems in which models autonomously sustain their own advancement.
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 · 2026-11-18
The authors' main claim, as the classifier picked it out 1.00
“Experiments with GLM-5.3 on PostTrainBench, a benchmark for autonomous post-training, show that training on RSI-Prime yields a 31.95% relative improvement in overall performance over the original GLM-5.3, highlighting a path toward future RSI systems in which models autonomously sustain their own advancement.”
How the main claim is stated: asserted 0.96
What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.08 · human participants 0.06 · controlled comparison 0.87 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.92 · mentions code or data availability 0.07
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.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Enabling large language models (LLMs) to autonomously conduct training and improve target models is an important foundation for recursive self-improvement (RSI). | background 1.00 | 0.06 | 0.05 | asserted 1.00 |
| As LLMs expand their role from being trained to conducting training, their model-training capability warrants dedicated study and systematic development. | background 0.68 | 0.06 | 0.03 | asserted 0.97 |
| To this end, we construct RSI-Prime, a dataset for improving model-training capability, comprising 350 complete experimental trajectories across seven task categories. | contribution 0.66 | 0.96 | 0.05 | asserted 0.99 |
| To help LLMs learn to make high-quality experimental decisions from training practice, we perform fine-grained decision identification and quality annotation, then apply decision-level loss masking to exclude supervision on low-quality behavior. | method 0.99 | 0.96 | 0.04 | asserted 0.91 |
| Experiments with GLM-5.3 on PostTrainBench, a benchmark for autonomous post-training, show that training on RSI-Prime yields a 31.95% relative improvement in overall performance over the original GLM-5.3, highlighting a path toward future RSI systems in which models autonomously sustain their own advancement. | result 0.98 | 0.96 | 0.84 | asserted 0.96 |
What kind of paper the classifier reports this is: dataset or benchmark 1.00 · evidence from observation or experiment 0.98
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, 2026-11-18: 1 request, 4,415 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 (4)
Figures are the authors', reproduced from the paper.
References (52)
- [3] Alfred V. Aho and Jeffrey D. Ullman , title =. 1972 on Pith
- [4] Publications Manual , year = "1983", publisher = on Pith
- [5] Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243 on Pithdoi:10.1145/322234.322243
- [6] Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of on Pith
- [7] Dan Gusfield , title =. 1997 on Pith
- [8] Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url = on Pith
- [9] Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages = on Pith
- [28] 2026 , month=sep, url= on Pith
- [31] Holistic Evaluation of Language Models , author =. Transactions on Machine Learning Research , year = on Pith
- [32] Wang, Yizhong and Mishra, Swaroop and Alipoormolabashi, Pegah and Kordi, Yeganeh and Mirzaei, Amirreza and Naik, Atharva and Ashok, Arjun and Dhanasekaran, Arut Selvan and Arunkumar, Anjana and Stap, David and Pathak, Eshaan and Karamanolakis, Giannis and Lai, Haizhi and Puroh... on Pith
- [35] A Self-Improving Coding Agent , author =. 2025 , eprint = on Pith
- [36] Self-Challenging Language Model Agents , author =. 2025 , eprint = on Pith
- [47] Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, Lilian Weng, and Aleksander Mądry. 2024. https://arxiv.org/abs/2410.07095 MLE-bench: Evaluating Machine Learning Agents on Machine Lear... on PithPith's reviewarXiv
- [48] Wanyi Chen, Xiao Yang, Xu Yang, Tianming Sha, Qizheng Li, Zhuo Wang, Bowen Xian, Fang Kong, Weiqing Liu, and Jiang Bian. 2026. https://arxiv.org/abs/2604.10547 Agent ^2 RL-Bench : Can LLM agents engineer agentic RL post-training? Preprint, arXiv:2604.10547 on PithPith's reviewarXiv
- [49] Zhixun Chen, Ming Li, Yuxuan Huang, Yali Du, Meng Fang, and Tianyi Zhou. 2025. https://doi.org/10.18653/v1/2025.findings-acl.1299 ATLAS : Agent tuning via learning critical steps . In Findings of the Association for Computational Linguistics: ACL 2025, pages 25334--25349, Vien... on Pithdoi:10.18653/v1/2025.findings-acl.1299
- [50] Yaxin Du, Xiyuan Yang, Zhifan Zhou, Yujie Ge, Cheng Wang, Jiajun Wang, Sijie Chen, Zehui Liu, Yuxin Zhang, Weicheng Gu, Julian Zhang, Zixing Lei, and Siheng Chen. 2026 a . https://arxiv.org/abs/2609.35561 RSI-Master : Structuring experiments to guide autonomous model improveme... on PithPith's reviewarXiv
- [51] Yaxin Du, Xiyuan Yang, Zhifan Zhou, Wanxu Liu, Zixing Lei, Zimeng Chen, Fenyi Liu, Haotian Wu, Yuzhu Cai, Zexi Liu, Xinyu Zhu, WenHao Wang, Linfeng Zhang, Chen Qian, and Siheng Chen. 2026 b . https://arxiv.org/abs/2605.10906 DataMaster: Data-Centric Autonomous AI Research . Pr... on PithPith's reviewarXiv
- [52] Haoyang Fang, Boran Han, Nick Erickson, Xiyuan Zhang, Su Zhou, Anirudh Dagar, Jiani Zhang, Ali Caner Turkmen, Cuixiong Hu, Huzefa Rangwala, Ying Nian Wu, Bernie Wang, and George Karypis. 2025. https://arxiv.org/abs/2505.13941 MLZero: A Multi-Agent System for End-to-end Machine... on PithPith's reviewarXiv
- [53] Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, and 8 others. 2026 a . https://arxiv.o... on PithPith's reviewarXiv
- [54] Shuzheng Gao, Zhaojian Yu, Penghao Yin, Yuling Shi, Kai Cai, Shilin He, and Michael R. Lyu. 2026 b . https://arxiv.org/abs/2611.03698 Training LLM agents to automate LLM post-training with RSI-Dojo . Preprint, arXiv:2611.03698 on PithPith's reviewarXiv
show 32 more
- [55] Shengran Hu, Cong Lu, and Jeff Clune. 2025. https://arxiv.org/abs/2408.08435 Automated Design of Agentic Systems . Preprint, arXiv:2408.08435 on PithPith's reviewarXiv
- [56] Qian Huang, Jian Vora, Percy Liang, and Jure Leskovec. 2023. https://arxiv.org/abs/2310.03302 MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation . arXiv preprint arXiv:2310.03302 on PithPith's reviewarXiv
- [57] Zhengyao Jiang, Dominik Schmidt, Dhruv Srikanth, Dixing Xu, Ian Kaplan, Deniss Jacenko, and Yuxiang Wu. 2025. https://arxiv.org/abs/2502.13138 AIDE: AI-Driven Exploration in the Space of Code . arXiv preprint arXiv:2502.13138 on PithPith's reviewarXiv
- [58] Feiyang Kang, Hanze Li, Adam Nguyen, Mahavir Dabas, Jiaqi W. Ma, Frederic Sala, Dawn Song, and Ruoxi Jia. 2026. https://arxiv.org/abs/2606.04261 Can Generalist Agents Automate Data Curation? Preprint, arXiv:2606.04261 on PithPith's reviewarXiv
- [59] Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, Christian Cosgrove, Christopher D. Manning, Christopher Re, Diana Acosta-Navas... on Pith
- [60] Joy Jia Yin Lim, Xin Huang, Hao Peng, Yaxi Lu, Xin Cong, Zhong Zhang, Maosong Sun, and Yankai Lin. 2026. https://arxiv.org/abs/2608.19072 What is Missing from AI Post-Training AI: An Empirical Analysis . Preprint, arXiv:2608.19072 on PithPith's reviewarXiv
- [61] Zerun Ma, Guoqiang Wang, Xinchen Xie, Yicheng Chen, He Du, Bowen Li, Yanan Sun, Wenran Liu, Kai Chen, and Yining Li. 2026. https://arxiv.org/abs/2604.14116 TREX: Automating LLM Fine-tuning via Agent-Driven Tree-based Exploration . Preprint, arXiv:2604.14116 on PithPith's reviewarXiv
- [62] Fanqing Meng, Lingxiao Du, Qiguang Chen, Ziqi Zhao, Haocheng Lu, Mengkang Hu, and Michael Qizhe Shieh. 2026. https://arxiv.org/abs/2607.25886 RSIBench-Data : Benchmarking data-centric research for recursive self-improvement . Preprint, arXiv:2607.25886 on PithPith's reviewarXiv
- [63] PostTrainBench Team . 2026 a . https://github.com/aisa-group/posttrainbench-website/blob/main/scores-v1.2.json PostTrainBench v1.2 Leaderboard Data . Accessed November 5, 2026 on Pith
- [64] PostTrainBench Team . 2026 b . https://posttrainbench.com/blog/posttrainbench-1-2/ PostTrainBench v1.2: More Reliable Scores and Verdicts on Pith
- [65] Ben Rank, Hardik Bhatnagar, Ameya Prabhu, Shira Eisenberg, Karina Nguyen, Matthias Bethge, and Maksym Andriushchenko. 2026. https://arxiv.org/abs/2603.08640 PostTrainBench : Can LLM agents automate LLM post-training? arXiv preprint arXiv:2603.08640 on PitharXiv
- [66] Maxime Robeyns, Martin Szummer, and Laurence Aitchison. 2025. https://arxiv.org/abs/2504.15228 A self-improving coding agent . Preprint, arXiv:2504.15228 on PithPith's reviewarXiv
- [67] Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. https://arxiv.org/abs/2303.11366 Reflexion: Language Agents with Verbal Reinforcement Learning . Preprint, arXiv:2303.11366 on PithPith's reviewarXiv
- [68] Patara Trirat, Wonyong Jeong, and Sung Ju Hwang. 2024. https://arxiv.org/abs/2410.02958 AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML . arXiv preprint arXiv:2410.02958 on PithPith's reviewarXiv
- [69] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. https://arxiv.org/abs/2305.16291 Voyager: An Open-Ended Embodied Agent with Large Language Models . Preprint, arXiv:2305.16291 on PithPith's reviewarXiv
- [70] Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, Eshaan Pathak, Giannis Karamanolakis, Haizhi Lai, Ishan Purohit, Ishani Mondal, Jacob Anderson, Kirby Kuzni... on Pithdoi:10.18653/v1/2022.emnlp-main.340
- [71] Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. 2025. https://arxiv.org/abs/2504.08066 The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search . arXiv preprint arXiv:2504.08066 on PithPith's reviewarXiv
- [72] Xingbo Yao, Xiaoman Wang, Zhengwu Lei, Tinghui Luo, YiLin Zhang, Yuefeng Wu, Yijie Xu, Tianfu Wang, Qingyuan Zhan, Ye Guo, Daoxin Zhang, Zhe Xu, Jian Liu, and Hui Xiong. 2026. https://arxiv.org/abs/2610.06347 ImproveAnyTask: An Autonomous Post-Training Harness for Iterative Mo... on PithPith's reviewarXiv
- [73] Zhaojian Yu, Penghao Yin, Shuzheng Gao, Shilin He, Kai Cai, and Xiao-Ping Zhang. 2026. https://arxiv.org/abs/2606.31551 AutoTrainess: Teaching Language Models to Improve Language Models Autonomously . Preprint, arXiv:2606.31551 on PithPith's reviewarXiv
- [74] Yunpeng Zhai, Shuchang Tao, Cheng Chen, Anni Zou, Ziqian Chen, Qingxu Fu, Shinji Mai, Li Yu, Jiaji Deng, Zouying Cao, Zhaoyang Liu, Bolin Ding, and Jingren Zhou. 2025. https://arxiv.org/abs/2511.10395 AgentEvolver: Towards Efficient Self-Evolving Agent System . Preprint, arXiv... on PitharXiv
- [75] Hangfan Zhang, Shao Zhang, Kangcong Li, Chen Zhang, Yang Chen, Yiqun Zhang, Lei Bai, and Shuyue Hu. 2026 a . https://arxiv.org/abs/2606.09498 Self-Harness: Harnesses That Improve Themselves . Preprint, arXiv:2606.09498 on PithPith's reviewarXiv
- [76] Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. 2026 b . https://arxiv.org/abs/2505.22954 Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents . Preprint, arXiv:2505.22954 on PithPith's reviewarXiv
- [77] Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, and Tatiana Shavrina. 2026 c . https://arxiv.org/abs/2603.19461 Hyperagents . Preprint, arXiv:2603.19461 on PithPith's reviewarXiv
- [78] Kai Zhang, Xiangchao Chen, Bo Liu, Tianci Xue, Zeyi Liao, Zhihan Liu, Xiyao Wang, Yuting Ning, Zhaorun Chen, Xiaohan Fu, Jian Xie, Yuxuan Sun, Boyu Gou, Qi Qi, Zihang Meng, Jianwei Yang, Ning Zhang, Xian Li, Ashish Shah, and 11 others. 2026 d . https://arxiv.org/abs/2510.08558... on PithPith's reviewarXiv
- [79] Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, Urmish Thakker, James Zou, and Kunle Olukotun. 2026 e . https://arxiv.org/abs/2510.04618 Agentic Context Engineering: Evolving Contexts for... on PithPith's reviewarXiv
- [80] Wenqi Zhang, Ke Tang, Hai Wu, Mengna Wang, Yongliang Shen, Guiyang Hou, Zeqi Tan, Peng Li, Yueting Zhuang, and Weiming Lu. 2024. https://arxiv.org/abs/2402.17574 Agent-Pro : Learning to evolve via policy-level reflection and optimization . Preprint, arXiv:2402.17574 on PithPith's reviewarXiv
- [81] Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. 2024. https://arxiv.org/abs/2308.10144 ExpeL: LLM Agents Are Experiential Learners . Preprint, arXiv:2308.10144 on PithPith's reviewarXiv
- [82] Zhengyang Zhao, Shengjie Ye, Lu Ma, Hao Liang, Hengyi Feng, and Wentao Zhang. 2026 a . https://arxiv.org/abs/2606.01279 ANDES: Agent Native Data Evolving Synthesis Tool for Autonomous Instruction Alignment . Preprint, arXiv:2606.01279 on PithPith's reviewarXiv
- [83] Ziqi Zhao, Fanqing Meng, Haocheng Lu, Lingxiao Du, Qiguang Chen, Mengkang Hu, and Xiao-Ming Wu. 2026 b . https://arxiv.org/abs/2609.36675 G \"o del forest: Balancing search depth and breadth for data-centric recursive self-improvement . Preprint, arXiv:2609.36675 on PithPith's reviewarXiv
- [84] Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Menglong Zhang, Yihang Chen, Jinsong Li, Runyu Yang, Qiangbin Liu, Xinlei Yu, Jianmin Zhou, Na Wang, Chunyang Sun, and Jun Wang. 2026. https://arxiv.org/abs/2603.18743 Memento-Skills : Let ag... on PitharXiv
- [85] Yifei Zhou, Sergey Levine, Jason Weston, Xian Li, and Sainbayar Sukhbaatar. 2025. https://arxiv.org/abs/2506.01716 Self-challenging language model agents . Preprint, arXiv:2506.01716 on PithPith's reviewarXiv
- [86] Adam Zweiger, Jyothish Pari, Han Guo, Ekin Akyürek, Yoon Kim, and Pulkit Agrawal. 2025. https://arxiv.org/abs/2506.10943 Self-Adapting Language Models . Preprint, arXiv:2506.10943 on PitharXiv
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.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment