Pith. sign in

REVIEW 4 major objections 4 cited by

FARS: A Fully Automated Research System Deployed at Scale

T0 review · 4 major / 0 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A fully automated multi-agent system can write complete AI research papers at scale, and a large human review of its live output maps both review-worthy work and recurring failure modes.

desk verdict Real systems paper: 166 uncurated AI-generated papers with full artifacts and a large human review study; the “occasionally strong” tail is thinner and noisier than the headline suggests. read the letter →

arxiv 2606.31651 v2 pith:T7S5LMXE submitted 2026-06-30 cs.AI

Qiong Tang , Tianxiang Sun , Xiangkun Hu , Xiangyang Liu , Yiran Chen , Yunfan Shao , Bobo Li , Changze Lv
show 49 more authors
Cheng Xu Chengsong Huang Chunyang Li Dizhan Xue Hao Bai Haodong Duan Hengquan Guo Hongyang He Hongyi Chen Hui Shen Jiahao Yuan Jiankai Sun Jikang Cheng Jinfeng Xu Jingqi Tong Jingye Chen Jinxiu Liu Jixuan Leng Junchi Yu Kaixun Jiang Kun Xiang Kunpeng Yao Lang Feng Liangqi Yuan Longsen Gao Meng Li Qi Jia Qiushi Sun Shengyuan Ding Shizhan Gong Siru Zhong Terry Jingchen Zhang Tianle Gu Tianyi Liang Weijie Liu Weikai Yang Weizhi Fei Xin Wang Xinpeng Liu Xuanwen Ding Yihong Tang Yuanli Wang Yukun Jiang Yuming Yang Zhengbao He Zhikai Chen Zhikun Xu Zhuang Li Zihao Huang
This is my paper
classification cs.AI
keywords automatedresearchmulti-agentsystemsAI-for-AIlarge-scaledeploymentevaluationintegrityauditlanguage-modelagentsscientificworkflowautomation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that automated research should be judged as a continuous deployment, not as a few hand-picked demos. It presents FARS, a system that turns open research directions into finished papers by running ideation, planning, experiments, and writing through specialized agents that share one auditable workspace of proposals, code, logs, results, and manuscripts. In a first public run, the system produced 166 complete AI and machine-learning papers across 67 fine-grained topics while keeping intermediate artifacts inspectable rather than hiding failures. Structured human reviews of most of that corpus indicate that some outputs are review-worthy and occasionally strong, while also exposing common limits: thin experimental evidence, methodological weaknesses, and integrity problems that only show up when claims are checked against code and logs. The sympathetic reading is that scalable idea-to-paper automation is already workable for AI-for-AI research, but scientific reliability still hinges on contribution, experimental sufficiency, and faithful writing.

What carries the argument

The shared auditable workspace linking four sequential stages—Ideation, Planning, Experiment, and Writing—so each project is a focused hypothesis-and-validation attempt whose proposals, plans, code, logs, results, and manuscript remain inspectable end to end.

What would settle it

Have independent expert reviewers, without self-selection and with mandatory artifact checks, re-score a random subset of the same papers and re-run key experiments from the released code and logs; if quality ranks collapse or claimed numbers routinely fail to match artifacts, the deployment claim does not hold.

Watch

Extended reading notes

Core claim

FARS shows that a fully automated, stage-structured research system can generate a large, unaudited-in-advance corpus of complete AI/ML papers and that expert structured review of that live distribution finds reviewable and sometimes strong artifacts alongside systematic shortfalls in experimental scope, methodology, and claim integrity.

Load-bearing premise

That volunteer reviewers who self-claim papers, rather than an independent venue committee, and who often use language models for polishing, give a fair enough reading of scientific quality and integrity for the whole deployment corpus.

Editorial extensions

If this is right

  • Automated research can be run as a continuous public pipeline that preserves weak and negative outcomes instead of only curated successes.
  • Human evaluation of AI-generated papers needs artifact-grounded integrity checks, not only surface scores or fluent presentation.
  • The main bottlenecks for reliable automated science are contribution value, experimental sufficiency, and claim faithfulness, not manuscript polish.
  • Progress means agents that design decisive experiments, audit claims against provenance, learn from failed trajectories, and withhold unsupported outputs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same deployment pattern will be harder in fields where hypotheses cannot be reduced to code, shared benchmarks, and GPU runs.
  • A permanent public artifact store of failed plans and negative results could become a training and auditing resource for later research agents.
  • If automated reviewers systematically rate the same corpus higher than expert humans, deployment monitoring cannot rest on automated review alone.
  • Treating short, falsifiable negative-result papers as first-class outputs may change how fields track progress under automation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 0 minor

Summary. The paper presents FARS, a multi-agent AI-for-AI research system that autonomously runs Ideation, Planning, Experiment, and Writing stages through a shared auditable workspace. In a continuous public deployment (417 hours, ~$186k, 21.6B tokens), FARS produced 166 complete papers spanning 67 AI/ML topics, preserving proposals, code, logs, results, and manuscripts rather than a curated success set. Quality is assessed via 282 structured volunteer reviews of 140 papers (ICLR-style overall ratings and sub-scores, integrity audit against artifacts, LLM-use disclosure) and a uniform Stanford Agentic Reviewer comparison against public outputs of six prior systems. The authors report a left-skewed quality distribution (paper-level mean overall rating 3.23; 11.4% of reviewed papers with mean ≥6), recurring weaknesses in experimental sufficiency and methodology, integrity flags on 27.9% of reviewed papers, and a higher automated SAR mean for FARS than for prior systems’ released papers.

Significance. If the deployment and evaluation hold, this is a substantial systems contribution: it shifts evidence for autonomous research from selected demos and human-scoped tasks to a live, large-scale idea-to-paper distribution with inspectable intermediate artifacts. Strengths that should be credited explicitly include the artifact-centered design (shared workspace, dual verification, checkpointed experiment loop), the decision to retain weak and negative trajectories, the AI Integrity Audit grounded in code/logs rather than text alone, public release of papers and most reviews, and unusually detailed reporting of cost, topic emergence, score distributions, agreement, and failure modes. For the AI-agent and automated-science communities, a calibrated account of what continuous end-to-end generation actually yields—and where it fails—is more useful than another curated showcase.

major comments (4)
  1. §5.2 and Figure 3: the load-bearing claim that FARS produces “occasionally strong” artifacts (abstract; §1; §7) rests on 16/140 papers with mean overall rating ≥6, of which 14 had a single review and only 2 of 95 multi-reviewed papers reached mean 6. Combined with self-claim assignment (§5.1) rather than blinded/centralized matching, the high-score prevalence is not yet established as a stable property of the corpus. Either (i) report multi-review-only statistics as the primary “strong tail” estimate, (ii) obtain additional independent reviews for the high-scoring singleton papers, or (iii) substantially soften “occasionally strong” to a claim about review-worthiness with a thin, sparsely multi-reviewed upper tail.
  2. §6 and Table 6: the cross-system ranking under SAR is presented as evidence that FARS is competitive with or stronger than prior systems, yet baseline n ranges from 3 to 71, release criteria differ, domains are non-matched, and SAR is systematically more lenient than the human panel on FARS (5.00 vs 3.23). The paper already labels this “for reference only,” but the abstract/conclusion still lean on competitiveness. Restrict Table 6 to a clearly secondary relative ranking, drop or heavily caveat percentage-at-threshold claims when no baseline paper reaches ≥6 under SAR, and do not use SAR to support scientific quality claims that the human study is meant to carry.
  3. §5.1–5.3: integrity and experimental-sufficiency findings are among the paper’s strongest results (flags concentrated in low ratings; experimental weaknesses in 92.9% of reviews; code-verifiable modes almost always escalated). However, the qualitative coding of strengths/weaknesses and integrity modes uses a single LLM annotator (Claude Opus 4.6, T=0) without reported human agreement or double-coding. Because these percentages structure the failure-mode narrative, provide inter-annotator agreement on a subsample or human validation of the category assignments; otherwise mark the category rates as exploratory.
  4. §3–4 vs evaluation: the system description asserts design choices intended to improve faithfulness (experiment contract, effectiveness gate, dual verification, blueprint-before-prose, numerical checks against artifacts), yet §5 still finds non-trivial integrity and sufficiency failures. The manuscript does not quantify how often gates fired, retries occurred, or negative-result paths were taken, nor ablate which stage mechanisms reduce integrity flags. Without process metrics linking mechanisms to outcomes, the architecture section overclaims causal contribution relative to the evaluation evidence. Add deployment process statistics (gate pass/fail rates, retries, negative-result fraction) or clearly separate design intent from measured effect.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: FARS quality claims rest on external volunteer human reviews and a third-party automated reviewer, not on self-defined success metrics or fitted-as-prediction loops.

full rationale

This is an empirical systems paper, not a first-principles derivation. The load-bearing claims are (i) that a continuous deployment produced 166 complete papers across 67 topics with preserved artifacts, and (ii) that 282 structured volunteer reviews of 140 papers (plus a uniform Stanford Agentic Reviewer comparison) show review-worthy and occasionally strong outputs alongside recurring experimental, methodological, and integrity failures. Neither claim reduces to its inputs by construction: paper counts and artifact traces are deployment facts; quality scores come from external human reviewers under an ICLR-style rubric and from an independent automated reviewer applied uniformly to FARS and prior systems’ public papers. There is no self-definitional X-from-Y loop, no parameter fitted to data and re-labeled as a prediction, no uniqueness theorem imported from the authors’ prior work to force the result, and no ansatz smuggled in via self-citation. Related-work citations to other autonomous research systems are ordinary literature positioning, not load-bearing premises of the evaluation. Author–reviewer name overlap (two Analemma authors appear in the named reviewer list) is a possible conflict-of-interest concern, not a circular reduction of a claimed derivation. Under the stated criteria the honest finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

As a systems-and-evaluation paper, FARS does not rest on fitted physical constants or new particles. Its load-bearing premises are engineering design choices (stage decomposition, experiment contracts, artifact workspace) and the assumption that volunteer structured reviews plus artifact checks are adequate proxies for scientific quality. Free parameters are operational (compute budget, seed topics, review form calibration). Invented entities are the named system and its stage agents.

free parameters (3)
  • seed research topics (9)
    Initial directions supplied by humans; Ideation is allowed to explore beyond them, but the starting distribution is a design choice that shapes the corpus.
  • compute and token budget per project
    Fixed budget and GPU cluster size constrain experiment scope; many reviewer complaints about narrow evaluation are partly consequences of this operational limit.
  • ICLR-style scoring rubric adapted for short focused papers
    Acceptance threshold of 6 and non-penalty for length are calibration choices that affect how many papers count as 'review-worthy'.
assumptions (4)
  • domain assumption A focused hypothesis plus an empirical or theoretical validation attempt is a valid unit of automated research, including well-documented negative results.
    Stated as a design principle in §3.1; underpins treating many short papers as successes rather than failures.
  • domain assumption Structured volunteer expert reviews with artifact access can measure soundness, contribution, and integrity of AI-generated papers.
    Core evaluation premise of §5; the paper itself notes the panel is not an independent venue PC.
  • ad hoc to paper Stage outputs exchanged through a shared workspace plus dual verification (agent self-assessment + harness checks) yield auditable experimental evidence.
    Architectural claim of the Experiment and Writing stages (§3.4–3.5).
  • domain assumption Standard multi-agent and ML tooling assumptions (LLM agents, GPU training stacks, curated skill libraries) behave as described under long-horizon unattended runs.
    Background engineering assumptions throughout §3.
invented entities (2)
  • FARS multi-agent pipeline (Ideation/Planning/Experiment/Writing + shared workspace) independent evidence
    purpose: Coordinate fully automated idea-to-paper research projects at scale while preserving intermediate artifacts.
    The system is the paper’s primary constructed object; independent evidence is the public deployment corpus and reviews, not an external physical measurement.
  • AI Integrity Audit (eight failure modes checked against code/logs) independent evidence
    purpose: Detect fabrication, hallucinated citations/methods, and design pathologies that ordinary peer review may miss for AI-generated papers.
    Review-protocol invention specific to this evaluation; evidence is the flag rates and free-text coding in §5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FARS: A Fully Automated Research System Deployed at Scale." pith.science (2026). https://pith.science/paper/T7S5LMXE

@misc{pith2026260631651,
  author       = {Pith},
  title        = {Pith review of: FARS: A Fully Automated Research System Deployed at Scale},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T7S5LMXE}},
  note         = {Machine review of arXiv:2606.31651}
}
read the original abstract

Recent automated research systems show that language-model agents can generate hypotheses, run experiments, and write complete manuscripts, but most evidence still comes from selected examples, human-framed topics, or a few pre-defined research tasks. We present FARS (Fully Automated Research System), a fully automated AI-for-AI research system designed to operate across research topics at scale. FARS autonomously generates and advances projects through ideation, planning, experimentation, and writing, using stage-specific agents coordinated through a shared workspace that records proposals, code, logs, results, and manuscripts. In its first public deployment, FARS produced 166 complete research papers spanning 67 fine-grained AI/ML topics while preserving intermediate artifacts as an auditable corpus rather than a curated set of successes. We evaluate this corpus with 282 structured reviews from volunteer reviewers covering 140 papers, including overall ratings, sub-scores, integrity checks, and LLM-use disclosure. The reviews indicate that FARS can produce review-worthy and occasionally strong AI/ML research artifacts in a large-scale public deployment, while also exposing recurring failure modes in narrow experimental scope, methodological limitations, and integrity issues.

Figures

Figures reproduced from arXiv: 2606.31651 by the authors.

Figure 1
Figure 1. Overview of the FARS architecture. The Ideation stage generates multiple research pro [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Distribution of human review scores across 282 reviews of 140 FARS-generated papers: [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Paper-level distribution of mean overall rating across the 140 reviewed papers, binned by [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Quality stratification and integrity findings for FARS papers. (a) Sub-score means by [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Topic distribution of the 166 FARS-generated papers. Blue bars indicate the nine seed [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning

    cs.AI 2026-08 conditional novelty 6.0 of 10

    Argus demonstrates that a fixed-weight, self-evolving multi-role agentic runtime with verification-gated persistence can achieve competitive benchmark results and retain reusable state across long-horizon tasks.

  2. IdeaTrail: Full-Process Agent Trajectories for Scientific Ideation

    cs.AI 2026-07 conditional novelty 6.0 of 10

    IdeaTrail provides 1,170 reverse-synthesized, multi-turn agent trajectories for scientific ideation, generated from known research artifacts via a Generator–Advisor loop with leakage and grounding checks.

  3. IdeaTrail: Full-Process Agent Trajectories for Scientific Ideation

    cs.AI 2026-07 conditional novelty 5.5 of 10

    IdeaTrail reverse-synthesizes 1,170 grounded multi-turn agent trajectories from real papers via a Generator–Advisor loop for scientific ideation process supervision.

  4. IdeaTrail: Full-Process Agent Trajectories for Scientific Ideation

    cs.AI 2026-07 unverdicted novelty 5.5 of 10

    IdeaTrail releases 1,170 reverse-to-forward Generator–Advisor trajectories that jointly log tools, evidence, intermediate artifacts, and reasoning for scientific ideation and proposal generation.

Reference graph

Works this paper leans on

26 extracted references · 19 linked inside Pith · cited by 2 Pith papers

  1. [1]

    org/abs/2411.14199

    URLhttps://arxiv. org/abs/2411.14199. Jinheon Baek, Sujay Kumar Jauhar, Silviu Cucerzan, and Sung Ju Hwang. Researchagent: Itera- tive research idea generation over scientific literature with large language models,

    arXiv

  2. [2]

    Alina Beygelzimer, Yann N

    URL https://arxiv.org/abs/2404.07738. Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan. Has the ma- chine learning review process become more arbitrary as the field has grown? the neurips 2021 consistency experiment,

    arXiv 2021

  3. [3]

    Hui Chen, Miao Xiong, Yujie Lu, Wei Han, Ailin Deng, Yufei He, Jiaying Wu, Yibo Li, Yue Liu, and Bryan Hooi

    URLhttps://arxiv.org/abs/2306.03262. Hui Chen, Miao Xiong, Yujie Lu, Wei Han, Ailin Deng, Yufei He, Jiaying Wu, Yibo Li, Yue Liu, and Bryan Hooi. Mlr-bench: Evaluating ai agents on open-ended machine learning research,

    arXiv

  4. [4]

    Ali Essam Ghareeb, Benjamin Chang, Ludovico Mitchener, Angela Yiu, Caralyn J

    URLhttps://arxiv.org/abs/2505.19955. Ali Essam Ghareeb, Benjamin Chang, Ludovico Mitchener, Angela Yiu, Caralyn J. Szostkiewicz, Dmytro Shved, Gavin J. Gyimesi, Jon M. Laurent, Samantha M. Wright, Muhammed T. Razzak, Andrew D. White, Silvia C. Finnemann, Michaela M. Hinks, and Samuel G. Rodriques. A multi-agent system for automating scientific discovery.Nature,

    arXiv

  5. [5]

    URLhttps://www.nature.com/articles/ s41586-026-10652-y

    doi: 10.1038/s41586-026-10652-y. URLhttps://www.nature.com/articles/ s41586-026-10652-y. Google. Nano Banana Pro: Gemini 3 Pro Image model from Google DeepMind.https://blog. google/innovation-and-ai/products/nano-banana-pro/, 11

    doi:10.1038/s41586-026-10652-y

  6. [6]

    Accessed: 2026-03-19. 16 Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Saizhuo Wang, Kun Zhang, Yuanzhuo Wang, Wen Gao, Lionel Ni, and Jian Guo. A survey on llm-as-a-judge,

    2026

  7. [7]

    Tarun Gupta and Danish Pruthi

    URLhttps://arxiv.org/abs/ 2411.15594. Tarun Gupta and Danish Pruthi. All that glitters is not novel: Plagiarism in AI generated research. InProceedings of ACL 2025,

    arXiv 2025

  8. [8]

    Intology

    URLhttps://arxiv.org/abs/ 2310.03302. Intology. Zochi technical report.arXiv,

    arXiv

Show all 26 references
  1. [9]

    Weld, and Peter Clark

    Peter Alexander Jansen, Oyvind Tafjord, Marissa Radensky, Pao Siangliulue, Tom Hope, Bhavana Dalvi, Bodhisattwa Prasad Majumder, Daniel S. Weld, and Peter Clark. CodeScientist: End- to-end semi-automated scientific discovery with code-based experimentation. InFindings of the A...

    2025

  2. [10]

    URLhttps://arxiv.org/abs/2511.15462. Patrick Tser Jern Kon, Jiachen Liu, Xinyi Zhu, Qiuyi Ding, Jingjia Peng, Jiarong Xing, Yibo Huang, Yiming Qiu, Jayanth Srinivasa, Myungjin Lee, Mosharaf Chowdhury, Matei Zaharia, and Ang Chen. Exp-bench: Can ai conduct ai research experiments?,

  3. [11]

    org/abs/2505.24785

    URLhttps://arxiv. org/abs/2505.24785. Ruochen Li, Teerth Patel, Qingyun Wang, and Xinya Du. Mlr-copilot: Autonomous machine learn- ing research based on large language models agents,

    arXiv

  4. [12]

    URLhttps://arxiv.org/abs/ 2408.14033. Yu Li, Chenyang Shao, Xinyang Liu, Ruotong Zhao, Peijie Liu, Hongyuan Su, Zhibin Chen, Qing- long Yang, Anjie Xu, Yi Fang, Qingbin Zeng, Tianxing Li, Jingbo Xu, Fengli Xu, Yong Li, and Tie-Yan Liu. AutoSOTA: An end-to-end automated researc...

  5. [13]

    URL https://arxiv.org/abs/2310.01783. Jiaqi Liu, Shi Qiu, Mairui Li, Bingzhou Li, Haonian Ji, Siwei Han, Xinyu Ye, Peng Xia, Zihan Dong, Meng Chen, Congyu Zhang, Letian Zhang, Guiming Chen, Haoqin Tu, Xinyu Yang, Lu Feng, Xujiang Zhao, Haifeng Chen, Jiawei Zhou, Xiao Wang, Wei...

    arXiv

  6. [14]

    Zijun Liu, Kaiming Liu, Yiqi Zhu, Xuanyu Lei, Zonghan Yang, Zhenhe Zhang, Peng Li, and Yang Liu

    URL https://arxiv.org/abs/2605.20025. Zijun Liu, Kaiming Liu, Yiqi Zhu, Xuanyu Lei, Zonghan Yang, Zhenhe Zhang, Peng Li, and Yang Liu. Aigs: Generating science from ai-powered automated falsification,

    arXiv

  7. [15]

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha

    URLhttps: //arxiv.org/abs/2411.11910. 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

  8. [16]

    URLhttps://arxiv.org/ abs/2408.06292. Rui Meng, Bhavana Dalvi Mishra, Jiefeng Chen, Chun-Liang Li, Palash Goyal, Mihir Parmar, Yiwen Song, Yale Song, Rajarishi Sinha, Parthasarathy Ranganathan, Burak Gokturk, Jinsung Yoon, and Tomas Pfister. Scientistone: Towards human-level a...

    arXiv

  9. [17]

    URLhttps://arxiv.org/abs/2605.26340. David B. Resnik, Mohammad Hosseini, and Rico Hauswald. Autonomous artificial intelligence, scientific research, and human values.AI and Ethics,

    arXiv

  10. [18]

    17 Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Michael Moor, Zicheng Liu, and Emad Barsoum

    doi: 10.1007/s43681-025-00908-0. 17 Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Michael Moor, Zicheng Liu, and Emad Barsoum. Agent laboratory: Using llm agents as research assistants,

    doi

  11. [19]

    Chenglei Si, Diyi Yang, and Tatsunori Hashimoto

    URLhttps://arxiv.org/abs/2501.04227. Chenglei Si, Diyi Yang, and Tatsunori Hashimoto. Can llms generate novel research ideas? a large-scale human study with 100+ nlp researchers,

    arXiv

  12. [20]

    Chenglei Si, Tatsunori Hashimoto, and Diyi Yang

    URLhttps://arxiv.org/abs/ 2409.04109. Chenglei Si, Tatsunori Hashimoto, and Diyi Yang. The ideation-execution gap: Execution outcomes of LLM-generated versus human research ideas.arXiv preprint arXiv:2506.20803,

    arXiv

  13. [21]

    Yixuan Weng, Minjun Zhu, Guangsheng Bao, Hongbo Zhang, Jindong Wang, Yue Zhang, and Linyi Yang

    URLhttps://arxiv.org/abs/2505.18705. Yixuan Weng, Minjun Zhu, Guangsheng Bao, Hongbo Zhang, Jindong Wang, Yue Zhang, and Linyi Yang. Cycleresearcher: Improving automated research via automated review. InThe Thirteenth International Conference on Learning Representations, 2025a...

    arXiv

  14. [22]

    Ruofeng Yang, Yongcan Li, and Shuai Li

    URLhttps://arxiv.org/abs/2504.08066. Ruofeng Yang, Yongcan Li, and Shuai Li. Aris: Autonomous research via adversarial multi-agent collaboration,

    arXiv

  15. [23]

    Jianxiang Yu, Zichen Ding, Jiaqi Tan, Kangyang Luo, Zhenmin Weng, Chenghua Gong, Long Zeng, Renjing Cui, Chengcheng Han, Qiushi Sun, Zhiyong Wu, Yunshi Lan, and Xiang Li

    URLhttps://arxiv.org/abs/2605.03042. Jianxiang Yu, Zichen Ding, Jiaqi Tan, Kangyang Luo, Zhenmin Weng, Chenghua Gong, Long Zeng, Renjing Cui, Chengcheng Han, Qiushi Sun, Zhiyong Wu, Yunshi Lan, and Xiang Li. Automated peer reviewing in paper sea: Standardization, evaluation, a...

    arXiv

  16. [24]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P

    URLhttps: //arxiv.org/abs/2407.12857. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena,

    arXiv

  17. [25]

    Kunlun Zhu, Jiaxun Zhang, Ziheng Qi, et al

    URLhttps://arxiv.org/ abs/2306.05685. Kunlun Zhu, Jiaxun Zhang, Ziheng Qi, et al. SafeScientist: Toward risk-aware scientific discoveries by LLM agents.arXiv preprint arXiv:2505.23559,

    arXiv

  18. [26]

    A DEPLOYMENTTOPICDISTRIBUTION Figure 5 groups the 166 generated papers by whether they fall under the nine seed topics provided at launch or under emergent topics discovered during autonomous exploration. B INTEGRITYFAILUREMODES The AI Integrity Audit in our review standard (S...

    arXiv

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.