Pith.

Artificial Intelligence · arXiv:2611.17790 v1 · submitted 13 Nov 2026 · 63 references

AutoRL: A Multi-Agent System towards Autonomous Reinforcement Learning

Simon Tang, Wenjie Shi, Yongjing Yin, Keheng Wang, Jialin Liu, Zheng Zhang, Linvo Lin, Xiangyuan Xue

show 3 more authors

Kan Ren, Jingang Wang, Xunliang Cai

The paper says

“Recursive self-improvement (RSI) is one of the most promising paths toward an AI singularity for superintelligence. Achieving this vision requires AI systems to undertake the full training pipeline for iterative improvement.”

more of the abstract

Reinforcement learning (RL) plays an essential role in this cycle to improve model capabilities and generalization, yet automating RL remains difficult due to unstable training dynamics and the need for sustained supervision. We introduce AutoRL, a multi-agent system for automating RL training, as a cornerstone for future RSI. Different from conducting research by end-to-end complete experiments, AutoRL monitors the evolving training state and integrates process-level supervision and research into an ongoing RL training, following the practical workflow of an RL researcher. Specifically, AutoRL consists of three RL-specific agentic roles. Perceptors track training dynamics, rollouts, evaluation, and efficiency, triggering deliberation when they detect anomalies, degradation, or stagnation. Three complementary Proposers diagnose problems and refine candidate interventions through shared evidence and checks. The Orchestrator reviews their recommendation and applies approved changes through online hot updates or checkpoint-based resumption. Experiments on Qwen3 and LongCat-Flash-Lite demonstrate the superiority of AutoRL over static RL recipes and a general agent. The research records show that existing agents can handle unattended RL supervision and dynamic adjustment. The difficulty for future RSI lies in developing fundamental methods that are valuable for resource-consuming validation.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 16 Nov 2026 · read the full text (116,076 characters) · not checked by a person

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

Holds up, with conditions.

10 minor

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

Are you an author? What you can do.

The paper, first character to last

The reviewer was given the whole text, 116,076 characters. Not checked: figures, code, what the 63 references say.

10 minor objections

The paper says, at §1 Introduction

Pith says: minor objection 1 of 10

The paper states 'Recursive self-improvement (RSI) is one of the most promising paths toward an AI singularity for superintelligence.' This is a strong claim that is not supported by evidence or citations. Consider rephrasing to reflect that this is a perspective held by some researchers.

The paper says, at §2.2 Specialized Roles and Shared Memory

Pith says: minor objection 2 of 10

The paper describes four Perceptors but Figure 1 shows four Perceptors while the text mentions 'Four Perceptors' in the abstract and 'Four Perceptors' in Section 2.2. However, in Section 3.1 Settings, the paper mentions 'AutoRL's supervisory agents use gpt-6.1-sol in both Qwen experiments and glm-5.2 in both LongCat-Flash-Lite experiments.' The number of Perceptors should be consistent throughout.

show 8 more

The paper says, at §3.1 Settings

Pith says: minor objection 3 of 10

The paper states 'Coding-agent evaluation uses SWE-related tasks.' It would be helpful to specify what SWE-related tasks are (e.g., a subset of SWE-bench) and how many tasks are used for evaluation.

The paper says, at §3.2 Main Results, Table 1

Pith says: minor objection 4 of 10

In Table 1, the numbers in parentheses for AutoRL show the difference compared to GRPO (Static Recipe). However, for Qwen3-1.7B, the AIME 2024 difference is listed as +3.33, but 30.00 - 26.67 = 3.33, which is correct. For AIME 2025, 30.00 - 21.67 = 8.33, correct. For HMMT 2025, 18.33 - 18.33 = 0.00, correct. For Beyond AIME, 12.50 - 11.00 = 1.50, correct. For IMO-AnswerBench, 14.63 - 8.63 = 6.00, correct. The calculations are consistent.

The paper says, at §3.4 Performance Plateau Breakthrough, Figure 3

Pith says: minor objection 5 of 10

The paper states 'AutoRL reaches 41.58% accuracy at step 180, exceeding its plateau maximum of 29.74% by 11.84%.' However, Figure 3 shows the plateau maximum around step 100-130, and the accuracy at step 180 is indeed higher. The paper should clarify whether the 29.74% is the maximum during the plateau or the accuracy at a specific step.

The paper says, at §4.2 Performance with Different LLMs, Figure 6

Pith says: minor objection 6 of 10

The paper compares GPT-6.1 and GLM-5.2 as backbones for AutoRL. It reports that GPT AutoRL achieves 41.58% peak accuracy while GLM AutoRL achieves 32.89%. However, the paper does not discuss the cost or computational requirements of the two backbones. If GLM-5.2 is significantly cheaper or faster, the trade-off may be acceptable. Consider adding a discussion of cost or efficiency.

The paper says, at §4.3 Scope and Limitations of AutoRL, Table 3

Pith says: minor objection 7 of 10

The paper states that 'Grading and data refinement account for 110 candidates (84.0%), making the correctness of the learning signal the dominant research focus.' This is informative, but the paper should also discuss whether this distribution reflects the actual needs of the training run or the biases of the LLM backend. For example, if the LLM is more likely to propose grading fixes than learning-rule changes, the distribution may not be optimal.

The paper says, at §5 Related Work

Pith says: minor objection 8 of 10

The paper cites several related works but does not provide a clear comparison of AutoRL with the most closely related systems (e.g., LLMZero, EvoTrainer, Athanor). Consider adding a table that summarizes the key differences in terms of intervention scope, experiment organization, and reuse of experience.

The paper says, at Appendix A.1 Task Configurations and Reproducibility

Pith says: minor objection 9 of 10

The paper states 'The 9,340 private LongCat-Flash-Lite mathematical reasoning problems mainly cover combinatorics and geometry, with additional problems in number theory, analysis, and probability.' Since the dataset is private, the paper should provide more details about its composition and how it was created to allow reproducibility.

The paper says, at Appendix A.6 Intervention, Recovery, and Failure Case Studies

Pith says: minor objection 10 of 10

The paper provides a detailed failure and recovery log in Table 5. However, the 'Recovery path' column includes entries like 'Checkpoint restart' and 'Agent: new job' without explaining what the agent did differently. For example, F3 says 'Agent: new job' but does not specify what the agent changed. Consider adding more detail about the agent's actions for each failure.

What would settle it

Repeat the Qwen3-4B campaign with a comparably capable open-weight model as the supervisory backbone; if AutoRL no longer beats static GRPO across the nine benchmarks, the gains belong to the proprietary LLM rather than the framework. A second decisive test is to remove the Reviewer Proposer and CPU-check stage: if accuracy stays unchanged, the deliberation machinery is not carrying the result.

from Pith's summary of the paper

How the review went

4 passes by deepseek-v4-flash · 16 Nov 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic concern withdrawn: Pith's reviewer misread the paper 6 s · 31,778 tokens read, 734 written
  3. Referee 10 minor 12 s · 32,055 tokens read, 2,864 written
  4. Check against the paper kept 0 of 6 objections 13 s · 33,323 tokens read, 3,112 written

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

Pith's passes raised 6 objections. A check against the paper's text kept 0 and withdrew 6 (5 the paper already says; 1 Pith's reviewer misread the paper).

The review in full

The stress test's concern

Withdrawn: Pith's reviewer misread the paper.

The paper says

“The general agent baseline uses native Codex~\citep{openai2026codexcli} with all built-in tools and features enabled, backed by the same \texttt{gpt-6.1-sol} model as the Qwen AutoRL runs.”

The referee's report

Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 5 of its 5 major objections did not survive that check.

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

The paper argues that the bottleneck for recursive self-improvement is not just whether an AI can train a model, but whether it can supervise and repair the unstable, failure-prone process of reinforcement learning while it runs. It proposes AutoRL, a multi-agent system that assigns separate specialized roles to monitoring a live RL experiment, diagnosing problems, and applying fixes either online or by resuming from checkpoints. The claimed payoff is that an ongoing RL run, not a finished model, becomes the object of autonomous research: the system decides when to intervene, how, and whether the intervention worked. If correct, this would move substantial everyday RL supervision work out of human hands and onto agents, while keeping the evaluation protocol fixed and the training budget unchanged.

What carries the argument. The load-bearing mechanism is the division of an RL researcher's job into three communicating role groups. Four Perceptors (Train, Rollout, Eval, Efficiency) watch complementary signals and issue a signal only when they detect anomalies, degradation, or stagnation, freezing their evidence into a shared batch. Three Proposers (Tuner, Innovator, Reviewer) each diagnose independently and then deliberate for at most five rounds, running CPU-only checks and sharing evidence before ranking candidates by Borda scoring and voting. The Orchestrator reviews the recommendation and applies it either as a HOT online update at a training-step boundary, retaining model and optimizer state, or as a COLD resumption from a compatible checkpoint, which is also the recovery path after a failed run. The HOT/COLD distinction is the device that lets research conclusions enter the training process without discarding accumulated progress.

Core claim. The central claim is that process-level supervision of a running RL experiment can be automated and yields better trained models than either a static recipe or a general-purpose coding agent. The authors report that on Qwen3-4B, AutoRL beats static-recipe GRPO on all nine benchmarks, with mathematical reasoning gains such as AIME 2024 rising from 46.67\% to 66.67\%, plus smaller gains on general knowledge and coding, and that it outperforms the Codex general agent on the same nine benchmarks. The paper further reports that in a 14.7-day LongCat-Flash-Lite coding-agent run, automated monitoring survived 11 runtime failures through checkpoint recovery and continued improving, with pass@1 on SWE-related tasks rising from 52.29\% to a peak of 65.13\%. The authors state the scope clearly: agents in this framework handle tuning, repair, and local reward refinement, but in their experiments produced no fundamental methodological breakthroughs, which they call a major unresolved challenge on the path to RSI.

The premise it rests on. The whole system rests on the closed-source supervising LLMs (GPT-6.1 Sol and GLM-5.2) being able to read training logs, diagnose the actual cause of a plateau or a NaN, and write a correct fix, so if those backends misdiagnose, AutoRL simply executes the wrong intervention more reliably than a human would.

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

Unattended long-horizon RL becomes feasible in principle: a run can survive repeated NaN, dtype, worker, and communication failures through checkpoint recovery, with learning continuing across interruptions.

Interventions that preserve model and optimizer state, via HOT updates, let later research build on everything learned so far rather than restarting from an earlier checkpoint.

Recipe repair and plateau investigation become automatable sub-tasks: fixing a flawed grader and adjusting entropy regularization produced renewed gains after prolonged stagnation in the reported runs.

The same supervisory workflow transfers across model scales and across LLM backbones, but the quality of what it achieves depends on which LLM supplies the diagnosis and the proposed fix.

As stated by the authors, the framework leaves a hard, unresolved boundary: agents refine an existing learning rule but do not, in these experiments, produce conceptually new learning methods.

If the supervising LLM is the binding constraint, AutoRL's ceiling rises with the base model's diagnostic skill; the framework's distinctive value may be converting a correct diagnosis into a validated, low-cost intervention rather than in generating the diagnosis itself.

Preserving optimizer state through HOT updates may matter more than per-benchmark accuracy: it means research effort compounds inside a single training run, a property that benchmark tables alone do not reveal.

A natural test the authors did not run is ablating the independent Reviewer and the CPU-check stage; if removing that scrutiny left accuracy unchanged, deliberation would be a costing exercise rather than a correctness filter.

The authors' closing remark that the hard part of RSI is 'developing fundamental methods that are valuable for resource-consuming validation' suggests the framework's next lever is cheaper validation, not better ideas.

Check this review

What was read

The passes were given all 116,076 characters of the paper's LaTeX source.

Fingerprint of that text (SHA-256): a86b6b3c80fbb2be6e438064ab804d72f78185c70781ae6cb3ff2bfcbf37aaa2

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 2611.17790

The passes, in the order they ran

  1. Reader deepseek-v4-flash · medium effort · $0.0036 · instructions of the Reader, fingerprint f9ead7300382…
  2. Skeptic deepseek-v4-flash · medium effort · $0.0030 · instructions of the Skeptic, fingerprint 8f88c0f61488…
  3. Referee deepseek-v4-flash · medium effort · $0.0034 · instructions of the Referee, fingerprint a852d99aea0d…
  4. Check against the paper deepseek-v4-flash · medium effort · $0.0036 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
  5. Circularity check deepseek-v4-flash · medium effort · $0.0029 · instructions of the Circularity check, fingerprint e9946e39d67b…
  6. Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0054 · instructions of the Summary for the page, fingerprint e4f56bf568e7…

The passes cost $0.022 in model calls together.

Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.ce10e2b95288

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 16 Nov 2026. The review you are reading still matches it.

Fingerprint of the review (SHA-256): 7ed99a12ce7d517791729311991d38fdc0f68a75351ca0a903d0647381144d6c

curl -s 'https://pith.science/api/v1/verdicts/2611.17790?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

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 · Review history (2 versions)

The paper

Abstract

Recursive self-improvement (RSI) is one of the most promising paths toward an AI singularity for superintelligence. Achieving this vision requires AI systems to undertake the full training pipeline for iterative improvement. Reinforcement learning (RL) plays an essential role in this cycle to improve model capabilities and generalization, yet automating RL remains difficult due to unstable training dynamics and the need for sustained supervision. We introduce AutoRL, a multi-agent system for automating RL training, as a cornerstone for future RSI. Different from conducting research by end-to-end complete experiments, AutoRL monitors the evolving training state and integrates process-level supervision and research into an ongoing RL training, following the practical workflow of an RL researcher. Specifically, AutoRL consists of three RL-specific agentic roles. Perceptors track training dynamics, rollouts, evaluation, and efficiency, triggering deliberation when they detect anomalies, degradation, or stagnation. Three complementary Proposers diagnose problems and refine candidate interventions through shared evidence and checks. The Orchestrator reviews their recommendation and applies approved changes through online hot updates or checkpoint-based resumption. Experiments on Qwen3 and LongCat-Flash-Lite demonstrate the superiority of AutoRL over static RL recipes and a general agent. The research records show that existing agents can handle unattended RL supervision and dynamic adjustment. The difficulty for future RSI lies in developing fundamental methods that are valuable for resource-consuming validation.

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 · 2026-11-16

The authors' main claim, as the classifier picked it out 0.88

“Experiments on Qwen3 and LongCat-Flash-Lite demonstrate the superiority of AutoRL over static RL recipes and a general agent.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.33 · human participants 0.03 · controlled comparison 0.87 · randomized assignment 0.02 · states quantitative result 0.05 · compares with prior results 0.92 · mentions code or data availability 0.03

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) is one of the most promising paths toward an AI singularity for superintelligence. background 1.00 0.04 0.04 asserted 0.83
Achieving this vision requires AI systems to undertake the full training pipeline for iterative improvement. background 0.99 0.05 0.03 asserted 1.00
Reinforcement learning (RL) plays an essential role in this cycle to improve model capabilities and generalization, yet automating RL remains difficult due to unstable training dynamics and the need for sustained supervision. background 0.53 0.05 0.03 asserted 1.00
We introduce AutoRL, a multi-agent system for automating RL training, as a cornerstone for future RSI. contribution 1.00 0.80 0.26 asserted 0.95
Different from conducting research by end-to-end complete experiments, AutoRL monitors the evolving training state and integrates process-level supervision and research into an ongoing RL training, following the practical workflow of an RL researcher. method 0.87 0.73 0.08 asserted 1.00
Specifically, AutoRL consists of three RL-specific agentic roles. method 0.94 0.76 0.03 asserted 1.00
Perceptors track training dynamics, rollouts, evaluation, and efficiency, triggering deliberation when they detect anomalies, degradation, or stagnation. method 0.99 0.71 0.03 asserted 1.00
Three complementary Proposers diagnose problems and refine candidate interventions through shared evidence and checks. method 0.99 0.70 0.04 asserted 0.99
The Orchestrator reviews their recommendation and applies approved changes through online hot updates or checkpoint-based resumption. method 0.99 0.66 0.05 asserted 1.00
Experiments on Qwen3 and LongCat-Flash-Lite demonstrate the superiority of AutoRL over static RL recipes and a general agent. result 1.00 0.95 0.60 asserted 1.00
The research records show that existing agents can handle unattended RL supervision and dynamic adjustment. result 0.74 0.24 0.09 asserted 0.99
The difficulty for future RSI lies in developing fundamental methods that are valuable for resource-consuming validation. future work 0.90 0.14 0.09 asserted 0.98

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99

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, 2026-11-16: 1 request, 8,743 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. AutoRL: autonomous RL supervision and research. Right: AutoRL continually adopts interventions, such as fixes and adjustments, to sustain gains over a static recipe. arXiv:2611.17790v1 [cs.AI] 13 Nov 2026
Figure 2 of the paper
Figure 2. Framework of AutoRL. Four Perceptors monitor training and diagnose problems. Three specialized Proposers discuss candidates, review evidence, and perform CPU-only checks to guide revision and selection. The Orchestrator reviews the recommendation, authorizes a HOT update or COLD checkpoint-based resumption, and confirms runtime activation. Shared records retain evidence, interventions, and checkpoint lineage. 2.2 Specialized Roles and Shared Memory To alleviate context pressure and concentrate attention on relevant evidence, AutoRL separates the training lifecycle into management,…
Figure 3 of the paper
Figure 3. Repairing the training recipe for LongCat-Flash-Lite. AutoRL improves reasoning accuracy as the GRPO (Static Recipe) run deteriorates. Light lines and markers show raw evaluations, and dark curves show a smooth average. Dashed lines mark six representative HOT updates and no restart occurs within the experiment. The star marks the highest raw evaluation accuracy. The upper axis gives elapsed wall-clock time. Effectiveness of recipe repair. As shown in
Figure 4 of the paper
Figure 4. Overcoming a performance plateau with Qwen3-4B. Shaded regions indicate initial improvement, the plateau, and renewed improvement. Light lines and markers show raw evaluations, and dark curves show a smooth average. Dashed lines mark applied interventions. Hollow diamonds mark COLD resumptions from checkpoints 100 and 135, with the modifications taking effect at steps 101 and 136. The star marks the highest raw evaluation accuracy. The upper axis gives elapsed wall-clock time. Improvement beyond the plateau. AutoRL reaches 41.58% accuracy at step 180, exceeding its plateau maximum…
Figure 5 of the paper
Figure 5. Recovering from failures during LongCat-Flash-Lite coding-agent training. The curve illustrates evaluations on SWE-related tasks, with a star marking the highest pass@1. Red dashed lines indicate failures. Paired crosses and dots connect the 11 failures to their recovery checkpoints, while the labels give the failure step and restored checkpoint. Two representative repairs are highlighted. The upper axis gives elapsed wall-clock time. Continued learning across failures. The run encounters 11 distinct failures, comprising seven NaN failures, one dtype mismatch, and three worker or…
Figure 6 of the paper
Figure 6. Comparison between AutoRL and general agent on Qwen3-4B. (a) Proposal composition and applied versus reverted interventions. (b) New and repeated training steps. (c) Raw accuracy (light), best-so-far accuracy (dark), and peaks (stars). Higher accuracy within the same step budget. As shown in
Figure 7 of the paper
Figure 7. AutoRL with different LLM backbones. (a) Online intervention shares over complete runs. (b) Distinct training steps reached. (c) Best accuracy (diamonds) and the mean of the last five evaluations (circles). To investigate the impact of using different LLMs as agent backends on the performance of AutoRL, we compare gpt-6.1-sol (GPT-6.1) and glm-5.2 (GLM-5.2) as AutoRL’s supervising LLM on the same Qwen3-4B task, retaining the GPT general agent as a reference. The runs share the initial model, training settings, and evaluation protocol

Figures are the authors', reproduced from the paper.

References (63)
  1. [1] Fang, Haoyang and Zhu, Wei and Han, Boran and Zhang, Alex and Pan, Zhenyu and Yang, Shuo and Zhang, Shuai and Gai, Jiading and Tang, Peng and Hu, Cuixiong and Zhu, Xuan and Rangwala, Huzefa and Karypis, George and Wang, Bernie , year =. 2606.18388 , archivePrefix = on PithPith's reviewarXiv
  2. [2] Chen, Guhong and Shi, Yingcheng and Li, Yongbin and Li, Binhua and Xu, Xander and Wei, Hu and Ni, Shiwen and Yang, Min and Ye, Jieping , year =. 2606.03108 , archivePrefix = on PithPith's reviewarXiv
  3. [3] He, Bingxiang and Sun, Pengqi and Peng, Shihao and Li, Shaoxuan and Xiao, Huiyu and Shi, Hongjian and Fu, Zixuan and Qiao, Ziqing and Zuo, Yuxin and Qian, Cheng and Yu, Xinqiang and Li, Jiayi and Chen, Xiusi and Song, Yizhen and Wang, Xiangyang and Lei, Iofai and Wang, Yudong ... on PithPith's reviewarXiv
  4. [4] Yamada, Yutaro and Lange, Robert Tjarko and Lu, Cong and Hu, Shengran and Lu, Chris and Foerster, Jakob and Clune, Jeff and Ha, David , year =. 2504.08066 , archivePrefix = on PithPith's reviewarXiv
  5. [5] Schmidgall, Samuel and Su, Yusheng and Wang, Ze and Sun, Ximeng and Wu, Jialian and Yu, Xiaodong and Liu, Jiang and Moor, Michael and Liu, Zicheng and Barsoum, Emad , year =. 2501.04227 , archivePrefix = on PithPith's reviewarXiv
  6. [6] Wu, Qingyun and Bansal, Gagan and Zhang, Jieyu and Wu, Yiran and Li, Beibin and Zhu, Erkang and Jiang, Li and Zhang, Xiaoyun and Zhang, Shaokun and Liu, Jiale and Awadallah, Ahmed Hassan and White, Ryen W and Burger, Doug and Wang, Chi , year =. 2308.08155 , archivePrefix = on PithPith's reviewarXiv
  7. [7] Hong, Sirui and Zhuge, Mingchen and Chen, Jiaqi and Zheng, Xiawu and Cheng, Yuheng and Zhang, Ceyao and Wang, Jinlin and Wang, Zili and Yau, Steven Ka Shing and Lin, Zijuan and Zhou, Liyang and Ran, Chenyu and Xiao, Lingfeng and Wu, Chenglin and Schmidhuber, Jürgen , year =. 2... on PithPith's reviewarXiv
  8. [8] Yilun Du and Shuang Li and Antonio Torralba and Joshua B. Tenenbaum and Igor Mordatch , title =. 2023 , howpublished =. 2305.14325 , archivePrefix = on PithPith's reviewarXiv
  9. [9] Noah Shinn and Federico Cassano and Edward Berman and Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao , title =. 2023 , howpublished =. 2303.11366 , archivePrefix = on PithPith's reviewarXiv
  10. [10] Sheng, Guangming and Zhang, Chi and Ye, Zilingfeng and Wu, Xibin and Zhang, Wang and Zhang, Ru and Peng, Yanghua and Lin, Haibin and Wu, Chuan , year =. 2409.19256 , archivePrefix = on PithPith's reviewarXiv
  11. [11] Fu, Wei and Gao, Jiaxuan and Shen, Xujie and Zhu, Chen and Mei, Zhiyu and He, Chuyi and Xu, Shusheng and Wei, Guo and Mei, Jun and Wang, Jiashu and Yang, Tongkai and Yuan, Binhang and Wu, Yi , year =. 2505.24298 , archivePrefix = on PithPith's reviewarXiv
  12. [12] Yang, Xu and Yang, Xiao and Fang, Shikai and Zhang, Yifei and Wang, Jian and Xian, Bowen and Li, Qizheng and Li, Jingyuan and Xu, Minrui and Li, Yuante and Pan, Haoran and Zhang, Yuge and Liu, Weiqing and Shen, Yelong and Chen, Weizhu and Bian, Jiang , year =. 2505.14738 , arc... on PitharXiv
  13. [13] Tang, Jiabin and Xia, Lianghao and Li, Zhonghang and Huang, Chao , year =. 2505.18705 , archivePrefix = on PithPith's reviewarXiv
  14. [14] Rank, Ben and Bhatnagar, Hardik and Prabhu, Ameya and Eisenberg, Shira and Nguyen, Karina and Bethge, Matthias and Andriushchenko, Maksym , year =. 2603.08640 , archivePrefix = on PitharXiv
  15. [15] Chen, Wanyi and Yang, Xiao and Yang, Xu and Sha, Tianming and Li, Qizheng and Wang, Zhuo and Xian, Bowen and Kong, Fang and Liu, Weiqing and Bian, Jiang , year =. 2604.10547 , archivePrefix = on PithPith's reviewarXiv
  16. [16] Fu, Qingxu and Liu, Boyin and Tao, Shuchang and Liu, Zhaoyang and Chen, Cheng and Jin, Xuanfa and Zhu, Rong and Ding, Bolin , year =. 2606.04484 , archivePrefix = on PithPith's reviewarXiv
  17. [17] Lee, Hyunji and Chen, Justin Chih-Yao and Singh, Joykirat and Khan, Zaid and Stengel-Eskin, Elias and Bansal, Mohit , year =. 2605.18565 , archivePrefix = on PithPith's reviewarXiv
  18. [18] Yao, Xingbo and Wang, Xiaoman and Lei, Zhengwu and Luo, Tinghui and Zhang, YiLin and Wu, Yuefeng and Xu, Yijie and Wang, Tianfu and Zhan, Qingyuan and Guo, Ye and Zhang, Daoxin and Xu, Zhe and Liu, Jian and Xiong, Hui , year =. 2610.06347 , archivePrefix = on PithPith's reviewarXiv
  19. [19] Guo, Taicheng and Chawla, Nitesh V. and Wiest, Olaf and Zhang, Xiangliang , year =. 2605.11518 , archivePrefix = on PithPith's reviewarXiv
  20. [20] Zhou, Yufa and Wang, Yichen and Fan, Jiakun and Shu, Ziqi and Liu, Zhichao and Zhang, Anru R. and Ling, Neiwen and Wei, Zijun , year =. 2610.21055 , archivePrefix = on PithPith's reviewarXiv
show 43 more
  1. [21] Shi, Zhan and He, Bing and Sang, Yisi and Dumoulin, Benoit and Lu, Hanqing , year =. 2606.20657 , archivePrefix = on PithPith's reviewarXiv
  2. [22] Meng, Fanqing and Du, Lingxiao and Lu, Haocheng and Chen, Qiguang and Zhao, Ziqi and Wu, Zijian and Zhuo, Jiayuan and Hu, Mengkang and Shieh, Michael Qizhe , year =. 2610.10310 , archivePrefix = on PithPith's reviewarXiv
  3. [23] Nathani, Deepak and Madaan, Lovish and Roberts, Nicholas and Bashlykov, Nikolay and Menon, Ajay and Moens, Vincent and Budhiraja, Amar and Magka, Despoina and Vorotilov, Vladislav and Chaurasia, Gaurav and Hupkes, Dieuwke and Cabral, Ricardo Silveira and Shavrina, Tatiana and ... on PithPith's reviewarXiv
  4. [24] Lupidi, Alisia and Gauri, Bhavul and Foster, Thomas Simon and Omari, Bassel Al and Magka, Despoina and Pepe, Alberto and Audran-Reiss, Alexis and Aghamelu, Muna and Baldwin, Nicolas and Cipolina-Kun, Lucia and Gagnon-Audet, Jean-Christophe and Leow, Chee Hau and Lefdal, Sandra... on PitharXiv
  5. [25] Du, Yaxin and Yang, Xiyuan and Zhou, Zhifan and Ge, Yujie and Wang, Cheng and Wang, Jiajun and Chen, Sijie and Liu, Zehui and Zhang, Yuxin and Gu, Weicheng and Zhang, Julian and Lei, Zixing and Chen, Siheng , year =. 2609.35561 , archivePrefix = on PithPith's reviewarXiv
  6. [26] Xu, Wujiang and Liang, Zujie and Mei, Kai and Gao, Hang and Tan, Juntao and Zhang, Yongfeng , year =. 2502.12110 , archivePrefix = on PithPith's reviewarXiv
  7. [27] Zhang, Qizheng and Hu, Changran and Upasani, Shubhangi and Ma, Boyuan and Hong, Fenglu and Kamanuru, Vamsidhar and Rainton, Jay and Wu, Chen and Ji, Mengmeng and Li, Hanchen and Thakker, Urmish and Zou, James and Olukotun, Kunle , year =. 2510.04618 , archivePrefix = on PithPith's reviewarXiv
  8. [28] Zhang, Yifan and Zou, Yunheng and Zhang, Shaokun and Hu, Jian and Zhang, Hao and Xu, Binfeng and Kautz, Jan and Dong, Yi , year =. 2609.18094 , archivePrefix = on PithPith's reviewarXiv
  9. [29] Zhao, Zihao and Zhang, Tunyu and Shi, Haizhou and Zhao, Yusong and Zhang, Xinxi and Wang, Hao , year =. 2610.04686 , archivePrefix = on PithPith's reviewarXiv
  10. [30] Bertalanič, Blaž and Fortuna, Carolina , year =. 2605.00914 , archivePrefix = on PithPith's reviewarXiv
  11. [31] Schulman, John and Wolski, Filip and Dhariwal, Prafulla and Radford, Alec and Klimov, Oleg , title =. 2017 , howpublished =. 1707.06347 , archivePrefix = on PithPith's reviewarXiv
  12. [32] Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , title =. 2024 , howpublished =. 2402.03300 , archivePrefix = on PithPith's reviewarXiv
  13. [33] Ouyang, Long and Wu, Jeff and Jiang, Xu and Almeida, Diogo and Wainwright, Carroll L. and Mishkin, Pamela and Zhang, Chong and Agarwal, Sandhini and Slama, Katarina and Ray, Alex and Schulman, John and Hilton, Jacob and Kelton, Fraser and Miller, Luke and Simens, Maddie and As... on PithPith's reviewarXiv
  14. [34] 2025 , howpublished =. 2501.12948 , archivePrefix = on PithPith's reviewarXiv
  15. [35] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , title =. 2022 , howpublished =. 2210.03629 , archivePrefix = on PithPith's reviewarXiv
  16. [36] Yang, John and Jimenez, Carlos E. and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , title =. 2024 , howpublished =. 2405.15793 , archivePrefix = on PithPith's reviewarXiv
  17. [37] 2024 , month = sep, howpublished = on Pith
  18. [38] 2026 , howpublished =. 2607.24653 , archivePrefix = on PitharXiv
  19. [39] 2026 , howpublished =. 2602.15763 , archivePrefix = on PithPith's reviewarXiv
  20. [40] 2026 , howpublished =. 2609.19969 , archivePrefix = on PitharXiv
  21. [41] Open-Source LLMs Collaboration Beats Closed-Source LLMs: A Scalable Multi-Agent System , author=. arXiv preprint arXiv:2507.14200 , year= on PithPith's reviewarXiv
  22. [42] Beyond Gemini-3-Pro: Revisiting LLM Routing and Aggregation at Scale , author=. arXiv preprint arXiv:2601.01330 , year= on PitharXiv
  23. [43] Yang, An and Li, Anfeng and Yang, Baosong and others , journal=. 2025 , url= on Pith
  24. [44] Yu, Qiying and Zhang, Zheng and Zhu, Ruofei and others , journal=. 2025 , url= on Pith
  25. [45] American Invitational Mathematics Examination , author=. 2024--2025 , howpublished= on Pith
  26. [46] 2025 , howpublished= on Pith
  27. [47] Towards Robust Mathematical Reasoning , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year= on Pith
  28. [48] Wang, Yubo and Ma, Xueguang and Zhang, Ge and others , journal=. 2024 , url= on Pith
  29. [49] Rein, David and Hou, Betty Li and Stickland, Asa Cooper and others , journal=. 2023 , url= on Pith
  30. [50] Program Synthesis with Large Language Models , author=. arXiv preprint arXiv:2108.07732 , year= on PithPith's reviewarXiv
  31. [51] Jain, Naman and Han, King and Gu, Alex and others , journal=. 2024 , url= on Pith
  32. [52] Liu, Zichen and Chen, Changyu and Li, Wenjun and others , journal=. Understanding. 2025 , url= on Pith
  33. [53] Scaling Embeddings Outperforms Scaling Experts in Language Models , author=. arXiv preprint arXiv:2601.21204 , year= on PitharXiv
  34. [54] Deng, Xiang and Da, Jeff and Pan, Edwin and others , journal=. 2025 , url= on Pith
  35. [55] 2026 , howpublished= on Pith
  36. [56] Loshchilov, Ilya and Hutter, Frank , title=. International Conference on Learning Representations , year= on Pith
  37. [57] Shoeybi, Mohammad and Patwary, Mostofa and Puri, Raul and others , title=. arXiv preprint arXiv:1909.08053 , year= on PithPith's reviewarXiv
  38. [58] Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and others , title=. Proceedings of the 29th Symposium on Operating Systems Principles , year= on Pith
  39. [59] Zwicker, William S. , title=. Handbook of Computational Social Choice , publisher=. 2016 , url= on Pith
  40. [60] Chen, Mingguang and Wang, Licheng and Qu, Bo , year =. 2607.07663 , archivePrefix = on PithPith's reviewarXiv
  41. [61] Zhang, Jenny and Hu, Shengran and Lu, Cong and Lange, Robert and Clune, Jeff , year =. 2505.22954 , archivePrefix = on PithPith's reviewarXiv
  42. [62] Chan, Alan and Winter, Christoph and Barto, Andrew and Pachocki, Jakub and Hinton, Geoffrey and Horvitz, Eric and Bengio, Yoshua and Song, Dawn and Clark, Jack and Greaves, Hilary and Korinek, Anton and Hammond, Samuel and Graepel, Thore and Bariach, Ben and Torr, Philip H. S.... on PithPith's reviewarXiv
  43. [63] Jafari, Akbar Anbar and Ozcinar, Cagri and Anbarjafari, Gholamreza , year =. 2511.10668 , archivePrefix = 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.

No comments yet.

Sign in to comment