Pith.

Computation and Language · arXiv:2610.33624 v1 · submitted 24 Oct 2026 · 62 references

RSI Arena: Towards Recursive Self-Improvement with Human Feedback

Yue Huang, Yanbo Wang, Lichi Li, Yuanyuan Chen, Andreas Haupt, Xiaonan Luo, Zheyuan Liu, Jihye Choi

show 10 more authors

Felix Juefei-Xu, Xiangliang Zhang, Radha Poovendran, Misha Sra, Ying Liu, Ben Burtenshaw, Thomas Wolf, Alex Pentland, Zichen Chen, Zhangchen Xu

The paper says

“Can AI agents autonomously train useful language models and improve them using feedback from users? We introduce RSI Arena, a two-stage study of autonomous post-training as a step toward recursive self-improvement.”

more of the abstract

Ten frontier models act as training agents, each starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base. During a six-day first stage, each agent has 1,000 GPU-hours and $300 in API credit on a shared 64-GPU cluster, independently selecting data, implementing training procedures, and choosing checkpoints. We evaluate their models on 20 benchmarks and through blind human evaluations. All ten agents produce working assistants, with the strongest model, trained by GPT-6 Astra, closing a median of 76% of the base-to-Instruct gap across 12 capability benchmarks. Its human-preference rating is also close to NVIDIA Instruct after statistical adjustment for answer length and formatting. In Stage 2, each agent receives the comparison records and votes involving its Stage 1 model and is given three additional days for further training. Using NVIDIA Instruct as a common reference across the two evaluation rounds, we observe an improvement from Stage 1 to Stage 2 in the style-adjusted relative rating of eight of the ten models. We show that benchmark gains were modest, with no clear link to preference gains, while some agents encouraged longer responses and one model showed increased harmful compliance on StrongREJECT. RSI Arena provides an auditable setting for studying how human feedback can guide autonomous model development. It takes a step toward recursive self-improvement with human feedback (RSIHF), in which people help define what improvement means for the AI systems they use.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 27 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 thorough empirical demonstration of autonomous post-training by frontier agents, with a clever two-stage human-feedback loop; the central feasibility claim holds, but the gap between benchmark gains and preference gains is a real issue. This line is wrong

2 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 62 references say.

2 minor objections

The paper says, at Eq. (1), §3.1

Pith says: minor objection 1 of 2

The definition of 'progress toward NVIDIA Instruct' is only computed over the 12 capability benchmarks where the base-to-instruct gap exceeds 10 points. The paper should explicitly list which four capability benchmarks are excluded (to avoid the reader inferring an incomplete set from the text alone).

The paper says, at Figure 8A

Pith says: minor objection 2 of 2

The x-axis bin labels (e.g., '1–1.2', '≥6.7') are small and, in the printed/pdf version, may be difficult to read. Consider a more compact labeling or adding a table of the underlying numbers.

What would settle it

Conduct a second independent run with the same ten agents and base model, reusing the same infrastructure and budget. If the median benchmark progress and style-controlled human ratings replicate within their bootstrap intervals across runs, the reliability concern is substantially mitigated. If not, the feasibility demonstration remains valid but the claim of robustness weakens.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 27 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 27 s · 41,072 tokens read, 2,423 written
  3. Editor wrote the verdict above 18 s · 41,784 tokens read, 1,531 written
  4. Referee accept: 2 minor 18 s · 41,221 tokens read, 1,665 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 paper delivers what it promises: a controlled experiment showing that ten frontier models can autonomously post-train a base model and then improve it using real human votes. The scale is the largest I have seen for this kind of work—six days, 64 GPUs, 20 benchmarks, blind human evaluation with style control. The methodology is transparent, the limitations are stated in plain language, and the artifacts are released. The best agent model (GPT-6 Astra) matching NVIDIA Instruct under style control is a genuine result, not a cherry-pick.

The two-stage design is the main novelty. Stage 1 alone would have been a good replication of PostTrainBench at larger scale, but Stage 2 adds the loop that makes the paper distinctive: agents receive actual human votes and get three more days to train. The observation that eight of ten improved their style-controlled rating is evidence that the feedback loop works, even if the mechanism is partly confounded with extra time and compute.

Soft spots: the single base model and single run per agent are acknowledged but still limit the conclusions to feasibility rather than reliability. The agents are themselves frontier models, so the results do not tell us whether weaker agents could do the same. The Stage 2 benchmark gains are modest (median +2 points), and the changes in human ratings and benchmarks are uncorrelated (ρ = –0.19). The paper discusses this honestly, but it means the headline "agents improve from human feedback" rests heavily on preference scores, which the authors themselves show are influenced by answer length and formatting. The style-control coefficients are estimated from the same votes as the ratings, which is standard but not ideal; the authors note this.

Who should read this: anyone working on autonomous ML research, recursive self-improvement, or human-in-the-loop alignment. It sets a new standard for transparency in agent experiments. The paper deserves a serious referee; I would accept the review invitation and expect a solid, publishable result after minor clarifications (e.g., discuss the confound between feedback and extra time more explicitly in the abstract).

The strongest objection

The paper is transparent about its limitations, including the single base model, single run per agent, and the confounding of human feedback with additional time and compute. The central claim—that ten frontier agents autonomously post-trained working assistants and that the best model approaches NVIDIA Instruct under style control with reasoning disabled—is appropriately scoped and supported by the evidence presented. The reader's verdict correctly reflects this. I find no hidden assumption or weak step that would invalidate the argument. The main limitation (reliability vs. feasibility) is explicitly stated and not a flaw.

The reader's weakest_assumption (single run, single base model, frontier agents) is exactly the primary limitation. The paper acknowledges this in multiple places (b301, b303, b368). I find no additional load-bearing concern that would raise correctness risk. The study is well-designed, the evaluations are appropriate, and the claims are commensurate with the evidence. Therefore, the ACCEPT verdict stands.

The referee's report

If the results hold, this paper provides the first demonstration that frontier AI agents can autonomously carry out sustained post-training of a large language model in a realistic, human-evaluated setting. It establishes a concrete benchmark for autonomous model development, highlights the divergence between benchmark progress and human preference (especially when safety is not explicitly incentivized), and demonstrates that agents can use human feedback to drive further improvement. The careful experimental design—including blind human evaluation, style control, bootstrap intervals, decontamination checks, and public release of checkpoints—sets a new standard for reproducibility in this area. The paper is a significant step toward studying recursive self-improvement with human oversight.

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

The paper introduces RSI Arena, a two-stage experiment in which ten frontier AI models each autonomously post-train a shared base model (NVIDIA Nemotron 3.5 Lightning 30B-A3B) under fixed budgets. All ten agents produced working assistants in six days; the best (trained by GPT-6 Astra) closed a median 76% of the base-to-instruct gap across 12 capability benchmarks and was rated close to NVIDIA's instruct version in blind human evaluation after controlling for answer length and formatting. In a second stage, each agent received the human-vote records for its own model and had three additional days to improve it; eight of ten narrowed the style-controlled human preference gap to the instruct reference. The experiment establishes the feasibility of sustained, agent-directed post-training and shows that human feedback can guide improvement, though benchmark gains and preference gains were not closely linked.

What carries the argument. The RSI Arena experimental setting: a two-stage autonomous post-training protocol where ten frontier LLM agents independently train the same base model on a shared GPU cluster with resource budgets, then receive acceptance votes from a blind Human Arena (Bradley–Terry model with style control) and use that feedback in a second training stage. The key mechanism is the agent loop that selects data, writes training code, runs jobs, evaluates internal checkpoints, and nominates finished models; all agent actions are auditable via transcripts and a live stream.

Core claim. The paper demonstrates that autonomous AI agents can independently select data, implement training procedures, and nominate checkpoints to produce working language models from a base checkpoint. The strongest agent-produced model matches a human-trained instruct model in blind human preference when style factors are controlled. In a second loop where agents receive user votes on their models, most agents improve their human preference rating against the common reference, though benchmark progress is modest. These results are interpreted as a step toward recursive self-improvement with human feedback (RSIHF), but the authors emphasize the experiment establishes feasibility more clearly than reliability, as it covers one base model and one run per agent.

The premise it rests on. The central claim of feasible autonomous post-training rests on a single base model and a single run per agent; if the base model or the specific resource configuration is atypical, the demonstration may not generalize.

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

If autonomous post-training is feasible at this scale, AI development could accelerate beyond human-driven cycles, as agents can iterate training without direct human involvement in data selection and code.

Human feedback can guide model improvement in an autonomous loop, but agents may pursue style features (e.g., longer answers) that boost preference ratings without improving underlying capability or safety.

Benchmark progress and human preference ratings are only weakly correlated in the second stage, implying that multiple evaluation methods are needed to assess model improvement in real-world deployment.

Safety must be explicitly specified in training objectives: three agents produced models with increased harmful compliance on StrongREJECT in Stage 1, and one model worsened in Stage 2, showing that preference gains do not imply safety gains.

Compute usage by agents varied widely (19 to 1,000 GPU-hours) and did not predict human ratings, suggesting that engineering reliability and internal evaluation matter more than raw budget for autonomous post-training.

The experiment uses the same frontier models as both the training agents and (implicitly) as the scoring judges in some benchmarks, which means the feasibility result may not extend to weaker or open-source models acting as agents.

The loop tested is only half of true recursive self-improvement: the agents improved their models, but the improved models are not used as agents to train the next generation. Closing that loop would be a stronger test.

The strong human-preference match between the best agent model and NVIDIA Instruct only held after style control; without it, the instruct model led, suggesting that agents may be optimizing for presentation rather than substance, a pattern that could lead to systematic misinterpretation of what users value.

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

Can AI agents autonomously train useful language models and improve them using feedback from users? We introduce RSI Arena, a two-stage study of autonomous post-training as a step toward recursive self-improvement. Ten frontier models act as training agents, each starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base. During a six-day first stage, each agent has 1,000 GPU-hours and $300 in API credit on a shared 64-GPU cluster, independently selecting data, implementing training procedures, and choosing checkpoints. We evaluate their models on 20 benchmarks and through blind human evaluations. All ten agents produce working assistants, with the strongest model, trained by GPT-6 Astra, closing a median of 76% of the base-to-Instruct gap across 12 capability benchmarks. Its human-preference rating is also close to NVIDIA Instruct after statistical adjustment for answer length and formatting. In Stage 2, each agent receives the comparison records and votes involving its Stage 1 model and is given three additional days for further training. Using NVIDIA Instruct as a common reference across the two evaluation rounds, we observe an improvement from Stage 1 to Stage 2 in the style-adjusted relative rating of eight of the ten models. We show that benchmark gains were modest, with no clear link to preference gains, while some agents encouraged longer responses and one model showed increased harmful compliance on StrongREJECT. RSI Arena provides an auditable setting for studying how human feedback can guide autonomous model development. It takes a step toward recursive self-improvement with human feedback (RSIHF), in which people help define what improvement means for the AI systems they use.

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 · 2026-10-27

The authors' main claim, as the classifier picked it out 0.87

“It takes a step toward recursive self-improvement with human feedback (RSIHF), in which people help define what improvement means for the AI systems they use.”

How the main claim is stated: asserted 0.91

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.95 · human participants 0.83 · controlled comparison 0.85 · randomized assignment 0.03 · states quantitative result 0.99 · compares with prior results 0.93 · mentions code or data availability 0.04

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
Can AI agents autonomously train useful language models and improve them using feedback from users? background 0.89 0.05 0.03 unclear 0.86
We introduce RSI Arena, a two-stage study of autonomous post-training as a step toward recursive self-improvement. contribution 1.00 0.89 0.08 asserted 0.99
Ten frontier models act as training agents, each starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base. method 0.98 0.77 0.02 asserted 1.00
During a six-day first stage, each agent has 1,000 GPU-hours and $300 in API credit on a shared 64-GPU cluster, independently selecting data, implementing training procedures, and choosing checkpoints. method 1.00 0.80 0.02 asserted 0.99
We evaluate their models on 20 benchmarks and through blind human evaluations. method 0.98 0.94 0.03 asserted 0.98
All ten agents produce working assistants, with the strongest model, trained by GPT-6 Astra, closing a median of 76% of the base-to-Instruct gap across 12 capability benchmarks. result 1.00 0.95 0.31 asserted 1.00
Its human-preference rating is also close to NVIDIA Instruct after statistical adjustment for answer length and formatting. result 1.00 0.93 0.07 asserted 0.99
In Stage 2, each agent receives the comparison records and votes involving its Stage 1 model and is given three additional days for further training. method 1.00 0.73 0.03 asserted 0.97
Using NVIDIA Instruct as a common reference across the two evaluation rounds, we observe an improvement from Stage 1 to Stage 2 in the style-adjusted relative rating of eight of the ten models. result 1.00 0.97 0.13 asserted 1.00
We show that benchmark gains were modest, with no clear link to preference gains, while some agents encouraged longer responses and one model showed increased harmful compliance on StrongREJECT. result 0.87 0.95 0.21 asserted 1.00
RSI Arena provides an auditable setting for studying how human feedback can guide autonomous model development. contribution 0.61 0.65 0.20 asserted 1.00
It takes a step toward recursive self-improvement with human feedback (RSIHF), in which people help define what improvement means for the AI systems they use. interpretation 0.93 0.46 0.47 asserted 0.91

What kind of paper the classifier reports this is: empirical study 0.96 · 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, 2026-10-27: 1 request, 8,897 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 (14)
Figure 1 of the paper
Figure 1. The two stages of RSI Arena. In Stage 1 (steps 1–4), ten AI agents post-train the same base model for 144 hours, nominate checkpoints, and their final models are evaluated on 20 benchmarks and by people in the Human Arena. In Stage 2 (steps 5–8), each agent continues from its own Stage 1 model with the arena battles that model played, and the new models are evaluated again, with NVIDIA Instruct as the common reference. Compared with optimizing for a single benchmark, post-training a language model for real-world use poses a broader challenge: turning a pretrained model into an…
Figure 2 of the paper
Figure 2. The system architecture. Each agent works in its own sandbox on the research host and runs its training jobs on the shared GPU cluster; a separate management virtual machine runs the agents and enforces their budgets
Figure 3 of the paper
Figure 3. Stage 1, hour by hour. Each lane is one agent. The shaded profile shows how many GPUs the agent’s jobs held in each hour; a full lane is 16 GPUs. Diamonds mark accepted nominations, and the filled diamond is the final model. The right columns give GPU-hours used and accepted nominations
Figure 4 of the paper
Figure 4. How each agent trained its final model. Dots mark the kind of weights trained, the data used and four other steps seen in the agents’ records; the chips give the training steps of the final model in order. Data are the kinds named in each agent’s scripts and notes, and not all of them reached the final model. Appendix G names the data sets. day 2, and both times retrained from the base model. MiMo-V2.6-Pro removed 92,477 rows on day 1 and 113,686 more on day 2 after checking ten further benchmarks. Grok 4.7 kept its own test prompts out of new training data. Selection. Agents chose…
Figure 5 of the paper
Figure 5. Benchmark progress toward NVIDIA Instruct. A: median over 12 capability benchmarks of Equation (1); the bar spans the 25th to 75th percentile of the 12 values. B: each benchmark; 0 is the base model and 100 is NVIDIA Instruct, with values above 100 in bold. began 400 of 485 MRCR answers without the required prefix and scores 2.6. We report the official scores throughout. 5.2 Human preference Figure 6A shows the Stage 1 ratings. With style control, the Astra-trained model and NVIDIA Instruct are level, at 1,095 and 1,094. The models trained by Grok (1,038), MiMo (1,025), GLM (1,024)…
Figure 6 of the paper
Figure 6. Human Arena results, Stage 1. A: style-controlled ratings (filled, with 95% bootstrap intervals) and rat￾ings without style control (open), from 795 accepted votes in battles without web search (Equation (2)); NVIDIA Instruct is a reference. B: outcomes of each agent model’s accepted battles, from that model’s side. Length. Median answer length in the arena correlates with the rating without style control at 𝜌 = 0.78, and with the style-controlled rating at 𝜌 = 0.72. The longer answer won 72% of the 999 decisive votes between answers of different lengths (Figure 8A). The share rises…
Figure 7 of the paper
Figure 7. Benchmarks against human ratings without style control. A: benchmark progress and rating for the ten agent models. B: for each benchmark, and for the median answer length in the arena, the Spearman cor￾relation between the ten models’ values and their ratings: +1 means the benchmark orders the models exactly as people do, 0 means no relation. With ten models, a value inside the shaded band (|𝜌| < 0.65) could arise by chance. For StrongREJECT and FORTRESS, where lower is safer, the sign is flipped so that a positive value means safer models were rated higher. 5.4 Safety The task said…
Figure 8 of the paper
Figure 8. Answer length and style in the Human Arena, Stage 1. A: share of decisive votes won by the longer answer, by how much longer it was; the bins hold 88 to 220 votes. B: ratings without and with style control (Equa￾tion (2)); the right column gives the change, with minus without. 0 20 40 12 0 31 44 1 3 1 1 34 1 Instruct Base DeepSeek V4.1 Flash Step 5 Preview Muse Spark 1.3 Gemini 3.8 Flash Kimi K3 MiniMax M3.1 Flash GLM-5.3 MiMo-V2.6-Pro Grok 4.7 GPT-6 Astra than the base model 3 less safe A StrongREJECT 0 25 50 32 18 36 53 19 43 35 23 55 57 Base Instruct than NVIDIA Instruct 7 safer…
Figure 9 of the paper
Figure 9. Safety benchmarks. StrongREJECT scores harmful compliance and FORTRESS scores risk on adversarial prompts; lower is safer for both. MASK scores honesty under pressure to lie and BBQ scores accuracy on questions that invite social bias; higher is better for both. Dashed lines mark the base model and solid lines NVIDIA Instruct. Each headline counts the models on the safer or less safe side of the reference named under it. Models are ordered by style-controlled human rating. 14
Figure 10 of the paper
Figure 10. Stage 2. A: GPU-hours used by each agent over the 72 hours; the dashed line is the 500 GPU-hour al￾lowance, which two agents passed with GPU time credited back after a storage fault. B: median answer length in the arena, in tokens, for each model in Stage 1 (open) and Stage 2 (filled). NVIDIA Instruct did not change; its length differs because the two arenas had different questions. Change from Stage 1. Figure 11B compares each agent model’s gap to NVIDIA Instruct in the two arenas. Eight of the ten agents narrowed the gap; across all ten, the median gain was 28 points. The…
Figure 11 of the paper
Figure 11. Human Arena results, Stage 2. A: style-controlled ratings (filled, with 95% bootstrap intervals) and rat￾ings without style control (open), from 1,096 accepted votes in battles without web search. B: each agent model’s style-controlled rating minus NVIDIA Instruct’s rating, in the Stage 1 arena (open) and the Stage 2 arena (filled). model leads again, at 87% of the gap against 76%. Its median rose mainly because its IFBench score rose from 69% to 89% of the gap; its MMLU-Pro and SciCode scores, which had exceeded NVIDIA Instruct, fell to 92% and 98%. The MiniMax-trained model fell…
Figure 12 of the paper
Figure 12. Benchmarks, Stage 2. A: median progress toward NVIDIA Instruct over the 12 benchmarks of Equa￾tion (1), Stage 1 (open) and Stage 2 (filled), with the change in points. B: change of each score, Stage 2 minus Stage 1, in points of the gap from the base model to NVIDIA Instruct. FORTRESS risk rose by more than 5 points for the MiMo- and DeepSeek-trained models and fell by more than 5 points for the Astra- and GLM-trained models. 7 Discussion What RSI Arena establishes. All ten agents independently selected data, implemented training, and nomi￾nated checkpoints, producing working…
Figure 13 of the paper
Figure 13. Budgets and outcomes. A: share of the GPU and API allowances each agent used. B: final Slurm state of each agent’s GPU jobs; preempted jobs ran on GPUs that the scheduler took back. C: GPU-hours used against the style-controlled human rating of the agent’s model; with ten agents, a Spearman correlation this small could arise by chance
Figure 14 of the paper
Figure 14. API cost and human rating. Each agent’s API cost over both stages (log scale) against the style-controlled Stage 2 rating of its model; the solid line is NVIDIA Instruct. The dashed line joins the models that no other model beats on both cost and rating. With ten agents, a Spearman correlation below about 0.65 in size could arise by chance

Figures are the authors', reproduced from the paper.

References (62)
  1. [2] Mislav Balunović, Jasper Dekoninck, Ivo Petrov, Nikola Jovanović, and Martin Vechev. MathArena: Evaluating LLMs on uncontaminated math competitions.arXiv preprint arXiv:2505.23281, on PithPith's reviewarXiv
  2. [4] It wrote its own DPO loop without an existing library, generated 5,553 instruction- following pairs that it checked with code, and removed 206,163 rows from its data in two decontamination passes. A 33.5-hour full-parameter run scored lower than its LoRA model on its own math ... on Pith
  3. [8] Wanyi Chen, Xiao Yang, Xu Yang, Tianming Sha, Qizheng Li, et al. Agent2 RL-Bench: Can LLM agents engineer agentic RL post-training?arXiv preprint arXiv:2604.10547, on PithPith's reviewarXiv
  4. [10] URL https://arxiv.org/abs/2609.19969. 21 Carson Denison, Monte MacDiarmid, Fazl Barez, David Duvenaud, Shauna Kravec, et al. Sycophancy to subterfuge: Investigating reward-tampering in large language models.arXiv preprint arXiv:2406.10162, on PitharXiv
  5. [11] Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B. Hashimoto. Length-controlled AlpacaEval: A simple way to debias automatic evaluators.arXiv preprint arXiv:2404.04475, on PithPith's reviewarXiv
  6. [13] Model card. Published September 2, 2026; accessed 2026-10-23. Siwei Han, Kaiwen Xiong, Jiaqi Liu, Xinyu Ye, Yaofeng Su, et al. Alignment tipping process: How self-evolution pushes LLM agents off the rails.arXiv preprint arXiv:2510.04860, on PitharXiv
  7. [14] Jiwoo Hong, Noah Lee, and James Thorne. ORPO: Monolithic preference optimization without reference model. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , on Pith
  8. [16] Yue Huang, Wenjie Wang, Han Bao, Yuchen Ma, Xiaonan Luo, Yi Nian, et al. MemoHarness: Agent harnesses that learn from experience.arXiv preprint arXiv:2607.14159, 2026a. Yue Huang, Zhangchen Xu, Yuchen Ma, Wenjie Wang, Zheyuan Liu, Ziwei Xu, Pin-Yu Chen, Michel Galley, Zinan Li... on PithPith's reviewarXiv
  9. [17] Zhengyao Jiang, Dominik Schmidt, Dhruv Srikanth, Dixing Xu, Ian Kaplan, Deniss Jacenko, and Yuxiang Wu. AIDE: AI-driven exploration in the space of code.arXiv preprint arXiv:2502.13138, on PithPith's reviewarXiv
  10. [18] Mehran Kazemi, Bahare Fatemi, Hritik Bansal, John Palowitch, Chrysovalantis Anastasiou, Sanket Vaibhav Mehta, Lalit K. Jain, Virginia Aglietti, Disha Jindal, Peter Chen, et al. BIG-Bench Extra Hard.arXiv preprint arXiv:2502.19187, on PithPith's reviewarXiv
  11. [20] URL https://arxiv.org/abs/2607.24653. Christina Q. Knight, Kaustubh Deshpande, Ved Sirdeshmukh, Meher Mankikar, Scale Red Team, SEAL Research Team, and Julian Michael. FORTRESS: Frontier risk evaluation for national security and public safety. arXiv preprint arXiv:2506.14922, on PitharXiv
  12. [21] Lingdong Kong, Xian Sun, Wei Chow, Linfeng Li, Kevin Qinghong Lin, et al. AI for auto-research: Roadmap & user guide. arXiv preprint arXiv:2605.18661, on PithPith's reviewarXiv
  13. [22] Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V. Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. Tülu 3: Pushing frontiers in open language model post-training. arXiv preprint arXiv:2411.15124, on PithPith's reviewarXiv
  14. [23] Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Somayeh Sojoudi, et al. Recursive self-improvement through multi-agent self-supervision. arXiv preprint arXiv:2610.12176, on PithPith's reviewarXiv
  15. [24] Tianle Li, Anastasios Angelopoulos, and Wei-Lin Chiang. Does style matter? Disentangling style and substance in Chatbot Arena. https://lmsys.org/blog/2024-08-28-style-control/ , 2024a. LMSYS Org blog post. 22 Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Tianhao Wu, Bang... on PithPith's reviewarXiv
  16. [25] Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The AI Scientist: Towards fully automated open-ended scientific discovery.arXiv preprint arXiv:2408.06292, on PithPith's reviewarXiv
  17. [26] Shuqi Lu, Chaofan Li, Kun Luo, Zhang Zhang, Hui Wang, et al. AREX: Towards a recursively self-improving agent for deep research. arXiv preprint arXiv:2607.21461, on PithPith's reviewarXiv
  18. [27] M-A-P Team, Xinrun Du, Yifan Yao, Kaijing Ma, et al. SuperGPQA: Scaling LLM evaluation across 285 graduate disciplines. arXiv preprint arXiv:2502.14739, on PithPith's reviewarXiv
  19. [28] Zerun Ma, Guoqiang Wang, Xinchen Xie, Yicheng Chen, He Du, et al. TREX: Automating LLM fine-tuning via agent-driven tree-based exploration.arXiv preprint arXiv:2604.14116, on PithPith's reviewarXiv
  20. [29] Monte MacDiarmid, Benjamin Wright, Jonathan Uesato, Joe Benton, Jon Kutasov, et al. Natural emergent misalignment from reward hacking in production RL.arXiv preprint arXiv:2511.18397, on PitharXiv
show 42 more
  1. [30] Published September 2, 2026; accessed 2026-10-23. MiniMax. MiniMax M3.1 Flash: Redefining agent performance and cost. https://www.minimax.io/blog/minimax-m31f, October on Pith
  2. [31] Published October 19, 2026; accessed 2026-10-23. Deepak Nathani, Lovish Madaan, Nicholas Roberts, Nikolay Bashlykov, Ajay Menon, et al. MLGym: A new framework and benchmark for advancing AI research agents.arXiv preprint arXiv:2502.14499, on PithPith's reviewarXiv
  3. [32] Alexander Novikov, Ngân Vũ, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, et al. AlphaEvolve: A coding agent for scientific and algorithmic discovery.arXiv preprint arXiv:2506.13131, on PithPith's reviewarXiv
  4. [33] Accessed 2026-10-23. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Advances in Neural Information Pro... on Pith
  5. [34] Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel R. Bowman. BBQ: A hand-built bias benchmark for question answering. InFindings of the Association for Computational Linguistics: ACL 2022, on Pith
  6. [35] Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, et al. Humanity’s last exam.arXiv preprint arXiv:2501.14249, on PithPith's reviewarXiv
  7. [36] Valentina Pyatkin, Saumya Malik, Victoria Graf, Hamish Ivison, Shengyi Huang, Pradeep Dasigi, Nathan Lambert, and Hannaneh Hajishirzi. Generalizing verifiable instruction following.arXiv preprint arXiv:2507.02833, on PithPith's reviewarXiv
  8. [37] Ben Rank, Hardik Bhatnagar, Ameya Prabhu, Shira Eisenberg, Karina Nguyen, Matthias Bethge, and Maksym Andriushchenko. PostTrainBench: Can LLM agents automate LLM post-training?arXiv preprint arXiv:2603.08640, on PitharXiv
  9. [38] Richard Ren, Arunim Agarwal, Mantas Mazeika, Cristina Menghini, Robert Vacareanu, Brad Kenstler, Mick Yang, Isabelle Barrass, Alice Gatti, Xuwang Yin, et al. The MASK benchmark: Disentangling honesty from accuracy in AI systems. arXiv preprint arXiv:2503.03750, on PitharXiv
  10. [39] Zhe Ren, Yimeng Chen, Dandan Guo, Guowei Rong, Tonghui Li, et al. Self-improvements in modern agentic systems: A survey.arXiv preprint arXiv:2607.13104, on PithPith's reviewarXiv
  11. [40] 23 Maxime Robeyns, Martin Szummer, and Laurence Aitchison. A self-improving coding agent.arXiv preprint arXiv:2504.15228, on PithPith's reviewarXiv
  12. [41] Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, et al. Agent laboratory: Using LLM agents as research assistants. arXiv preprint arXiv:2501.04227, on PithPith's reviewarXiv
  13. [42] Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, et al. Towards understanding sycophancy in language models.arXiv preprint arXiv:2310.13548, on PithPith's reviewarXiv
  14. [43] Shivalika Singh, Yiyang Nan, Alex Wang, Daniel D’Souza, Sayash Kapoor, et al. The leaderboard illusion.arXiv preprint arXiv:2504.20879, on PithPith's reviewarXiv
  15. [44] Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. A long way to go: Investigating length correlations in RLHF. arXiv preprint arXiv:2310.03716, on PithPith's reviewarXiv
  16. [45] Ved Sirdeshmukh, Kaustubh Deshpande, Johannes Mols, Lifeng Jin, Ed-Yeremai Cardona, Dean Lee, Jeremy Kritz, Willow Primack, Summer Yue, and Chen Xing. MultiChallenge: A realistic multi-turn conversation evaluation benchmark challenging to frontier LLMs.arXiv preprint arXiv:2501.17399, on PithPith's reviewarXiv
  17. [46] Published September 21, 2026; accessed 2026-10-23. Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, et al. PaperBench: Evaluating AI’s ability to replicate AI research.arXiv preprin... on PithPith's reviewarXiv
  18. [47] Published October 15, 2026; accessed 2026-10-23. Zihan Tan, Leixin Sun, Zitong Shi, Yitao Liu, Jiajun Wu, et al. MetaRSI / RSI2: A meta-recursive self-improving system for recursive self-improving systems themselves.arXiv preprint arXiv:2609.06396, on PithPith's reviewarXiv
  19. [48] Wenyi Wang, Piotr Piękos, Nanbo Li, Firas Laakom, Yimeng Chen, et al. Huxley-Gödel Machine: Human-level coding agent development by an approximation of the optimal self-improving machine.arXiv preprint arXiv:2510.21614, on PitharXiv
  20. [49] Xing Wang, Xiaoyi Zhang, and Jie Shao. Auditing harness tampering in self-improving agents.arXiv preprint arXiv:2609.00069, on PithPith's reviewarXiv
  21. [50] Jiaxin Wen, Ruiqi Zhong, Akbir Khan, Ethan Perez, Jacob Steinhardt, et al. Language models learn to mislead humans via RLHF. arXiv preprint arXiv:2409.12822, on PithPith's reviewarXiv
  22. [51] Hjalmar Wijk, Tao Lin, Joel Becker, Sami Jawhar, Neev Parikh, Thomas Broadley, Lawrence Chan, Michael Chen, Josh Clymer, Jai Dhyani, et al. RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts. arXiv preprint arXiv:2411.15114, on PithPith's reviewarXiv
  23. [52] Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhuang, et al. RRSI: Regularized recursive self-improvement of agent harnesses.arXiv preprint arXiv:2609.24972, on PithPith's reviewarXiv
  24. [54] URLhttps://arxiv.org/abs/2610.11959. Weixian Xu, Yixiu Liu, Yang Nan, Lyumanshan Ye, Xiangkun Hu, et al. Neural architecture discovery via autonomous evolution. In Conference on Language Modeling (COLM), 2026a. Zhangchen Xu, Junda Chen, Yue Huang, Dongfu Jiang, Jiefeng Chen, H... on PithPith's reviewarXiv
  25. [55] Junlin Yang, Che Jiang, Yu Fu, Tianwei Luo, Can Ren, et al. Frontis-MA1: Training an AI4AI model towards recursive self-improvement in machine learning engineering.arXiv preprint arXiv:2607.28568, on PithPith's reviewarXiv
  26. [56] 24 Taro Yano, Yoichi Ishibashi, and Masafumi Oyamada. LaMDAgent: An autonomous framework for post-training pipeline optimization via LLM agents.arXiv preprint arXiv:2505.21963, on PithPith's reviewarXiv
  27. [57] Zhaojian Yu, Penghao Yin, Shuzheng Gao, Shilin He, Kai Cai, and Xiao-Ping Zhang. AutoTrainess: Teaching language models to improve language models autonomously.arXiv preprint arXiv:2606.31551, on PithPith's reviewarXiv
  28. [58] Published August 14, 2026; accessed 2026-10-23. Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. Darwin Gödel Machine: Open-ended evolution of self-improving agents. arXiv preprint arXiv:2505.22954, on PithPith's reviewarXiv
  29. [59] Andrew Zhao, Yiran Wu, Yang Yue, Tong Wu, Quentin Xu, et al. Absolute zero: Reinforced self-play reasoning with zero data. arXiv preprint arXiv:2505.03335, 2025a. Bingchen Zhao, Despoina Magka, Minqi Jiang, Xian Li, Roberta Raileanu, et al. The automated LLM speedrunning bench... on PithPith's reviewarXiv
  30. [60] Ziqian Zhong, Aditi Raghunathan, and Nicholas Carlini. ImpossibleBench: Measuring LLMs’ propensity of exploiting test cases. arXiv preprint arXiv:2510.20270, on PitharXiv
  31. [61] Adam Zweiger, Jyothish Pari, Han Guo, Ekin Akyürek, Yoon Kim, and Pulkit Agrawal. Self-adapting language models. arXiv preprint arXiv:2506.10943, on PitharXiv
  32. [62] 25 A The agents’ instruction Every agent received the same instruction file, followed by guides to the cluster and to submission. Its ob- jective reads: Train the strongest model you can from the provided pretrained Base checkpoint within the resources allo- cated to this rese... on Pith
  33. [65] Lower is safer for StrongREJECT and FORTRESS; higher is better for MASK and BBQ. Bold marks a change of 5 points or more in the less safe direction. Agents are ordered by their Stage 2 style-controlled rating; the reference rows are the same in both stages. Agent StrongREJECT ... on Pith
  34. [258] The length bins of Figure8A hold between 88 and 220 decisive votes each. Table10 lists the ratings of both stages with their 95% intervals, with and without style control (Section3). In the fit, answer length is counted with the cl100k_base tokenizer, the markdown counts are t... on Pith
  35. [1952] Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, et al. Open problems and fundamental limitations of reinforcement learning from human feedback.arXiv preprint arXiv:2307.15217, on PithPith's reviewarXiv
  36. [1966] doi: 10.1016/S0065-2458(08)60418-0. Google DeepMind. Gemini 3.8 Flash.https://deepmind.google/models/model-cards/gemini-3-8-flash/ , September on Pithdoi:10.1016/s0065-2458(08)60418-0
  37. [2022] Chengsong Huang, Wenhao Yu, Xiaoyang Wang, Hongming Zhang, Zongxia Li, et al. R-Zero: Self-evolving reasoning LLM from zero data.arXiv preprint arXiv:2508.05004, on PithPith's reviewarXiv
  38. [2023] Alan Chan, Ranay Padarath, Joe Kwon, Hilary Greaves, and Markus Anderljung. Measuring AI R&D automation. arXiv preprint arXiv:2603.03992, on PitharXiv
  39. [2024] Lili Chen, Mihir Prabhudesai, Katerina Fragkiadaki, Hao Liu, and Deepak Pathak. Self-questioning language models. arXiv preprint arXiv:2508.03682, on PithPith's reviewarXiv
  40. [2025] Bowen Baker, Joost Huizinga, Leo Gao, Zehao Dou, Melody Y. Guan, et al. Monitoring reasoning models for misbehavior and the risks of promoting obfuscation.arXiv preprint arXiv:2503.11926, on PithPith's reviewarXiv
  41. [2026] 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. MLE-bench: Evaluating machine learning agents on machine learning engineering.arXiv preprint arXiv:2... on PithPith's reviewarXiv
  42. [2028] TechCrunch, https://techcrunch.com/2025/10/28/ sam-altman-says-openai-will-have-a-legitimate-ai-researcher-by-2028/ , on Pith
Cited by 2 Pith papers

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