Pith. sign in

REVIEW 4 major objections 4 minor 6 cited by

Physics Supernova, an AI agent with physics-specific tools, reports a 23.5/30 score on the 2025 International Physics Olympiad theory problems, ranking 14th among 406 human contestants and above the human gold-medal median.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

An LLM agent with image-reading and answer-review tools scored 23.5/30 on IPhO 2025 theory problems, matching the median gold-medalist theory score.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection First agentic run on fresh IPhO 2025 theory problems, but 'surpasses gold-medal median' rests on a 0.7-point gap inside the run-to-run noise and self-performed grading. the 4 major comments →

arxiv 2509.01659 v1 pith:4DSWE7FX submitted 2025-09-01 cs.AI

Physics Supernova: AI Agent Matches Elite Gold Medalists at IPhO 2025

Jiahao Qiu , Jingzhe Shi , Xinzhe Juan , Zelin Zhao , Jiayi Geng , Shilong Liu , Hongru Wang , Sanfeng Wu
show 1 more author
Mengdi Wang
This is my paper
classification cs.AI
keywords physics olympiadAI agentstool integrationlarge language modelsphysics problem solvingIPhO 2025image measurementanswer review
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper claims that a tool-equipped AI agent, Physics Supernova, can solve International Physics Olympiad theory problems at the level of elite human gold medalists. Scored against the official 2025 IPhO marking scheme, the agent's five-run mean is 23.5/30, which would rank it 14th of 406 human contestants and put it above the 22.8 median theory score of the gold medalists. The point is not that a single model is smarter; it is that a flexible agent loop plus two physics-oriented tools—one for reading measurements from figures and one for reviewing its own answers—raises a strong base LLM to Olympiad-level performance. The paper supports this with ablations showing each tool adds points, with the largest gains on the hardest problems. If true, this shifts the practical question in scientific AI from raw model capability to principled tool integration.

Core claim

The paper reports that Physics Supernova matches elite human performance on the theory papers of IPhO 2025: a five-run mean of 23.5/30 (SD 0.8), corresponding to rank 14 among 406 human contestants and exceeding the median theory score of 22.8 for human gold medalists. The system is an agent rather than a single model: a manager LLM iterates a reason-act loop, calls an ImageAnalyzer tool to make precise measurements from problem figures, and calls an AnswerReviewer tool to check its own solutions for errors. The tool contributions are quantified by ablation: removing the reviewer drops the mean total from 23.5 to 22.4, removing the image tool drops it to 22.8, and using the base LLM alone gi

What carries the argument

The load-bearing mechanism is the manager agent's reason-act loop plus two dedicated tools. The ImageAnalyzer routes a high-resolution figure to a vision-language model and returns precise measurements, reducing mean absolute error in figure readings from 0.015 to 0.004 in the part-C case study. The AnswerReviewer prompts a second LLM to act as an uncompromising peer reviewer, checking dimensional consistency, signs, numerical mistakes, and missing steps, and reminding the worker to use the image tool for measurements. The manager is deliberately not given a fixed workflow; it chooses which tool to call as it builds up a trajectory of facts, assumptions, and partial results. The ablation num

Load-bearing premise

The result rests on the authors' own grading of the agent's output, since raw answers and grading sheets are not released; any unintended leniency in that grading would undo the headline score and rank.

What would settle it

Have two independent official IPhO graders, who have not seen the paper's runs, rescore the agent's raw answer sheets for all five runs under the official IPhO 2025 rubric; if the independently confirmed mean falls below the gold-medal median of 22.8 (or if raw answers are unavailable), the headline claim collapses. A cheaper alternative: run the same system on IPhO 2026 theory problems immediately after release and compare against contestants under identical scoring.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On the hardest 2025 theory problem (Theory 2), the agent's 6.08 mean exceeds the human top-10% score of about 5.0, while on the easiest problem its margin over the base LLM is small; difficulty is where the agent loop pays off.
  • Removing the answer-review tool drops the total from 23.5 to 22.4, and removing the image-reading tool drops it to 22.8, showing both tools are load-bearing.
  • The image-reading tool cuts measurement mean absolute error from 0.015 to 0.004 on figure-based readings, which means accurate visual extraction is a real component of physics reasoning, not a side detail.
  • Adding a computational knowledge engine lifts 3-digit-accurate answers on expert-knowledge queries from 3/10 to 9/10, indicating the same architecture can absorb more domain-specific tools.
  • Because the IPhO 2025 problems were released in July 2025 and use part-level rubrics, they offer a substantially fresher and finer-grained evaluation than older answer-only olympiad benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A blind, independent rescoring of the five runs is the obvious next experiment; without released raw answers, the rank-14 headline should be read as an in-house measurement.
  • The same agent should be run on an announced-but-unpublished olympiad (e.g., IPhO 2026) under identical tool settings; that would separate genuine transfer from any residual familiarity with the 2025 problems.
  • The ablation pattern suggests that weaker base LLMs may recover much of the gap when given the same tools; that is a cheap, testable way to map how much of the result is tool architecture versus frontier-model capability.
  • If the result transfers, physics-agent performance may be better improved by building better perception and verification tools than by scaling physics pretraining alone; analogous tool sets could then be applied to other quantitative sciences.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. This paper introduces Physics Supernova, an agent system built on the smolagents CodeAgent framework with Gemini 2.5 Pro as the manager, augmented with two task-specific tools: an ImageAnalyzer (a vision-language model) and an AnswerReviewer (an LLM-based peer-reviewer). The authors evaluate the system on the three theory problems of the 2025 International Physics Olympiad (IPhO), scoring the agent's outputs using official marking schemes applied by human experts. They report a mean theory score of 23.5/30 over five runs (SD 0.8), which they state ranks 14th among 406 contestants and exceeds the gold-medal median of 22.8. They also report ablations removing each tool, a case study on image reading, and an auxiliary expert-knowledge task set with a WolframAlpha tool.

Significance. If substantiated, this would be a notable milestone: a general LLM-based agent matching elite human performance on a fresh, fine-grained physics benchmark. The benchmark itself is strong—official IPhO 2025 problems, official human standings, and external official marking criteria from a third-party site. The paper also ships code and detailed prompts, supporting reproducibility of the method. However, the headline comparison currently rests on only five runs, with a mean advantage smaller than the run-to-run standard deviation, and the grading is not independently audited. The significance is therefore conditional on additional statistical evidence and measurement transparency.

major comments (4)
  1. [§3.2, Table 2] The claim that Physics Supernova 'surpasses the median gold-medal theory score' rests on a 0.7-point gap (23.5 vs. 22.8) with an SD of 0.8 over 5 runs. No significance test or confidence interval is reported, and per-run total scores and per-run ranks are not provided. With these numbers, a one-sample t-test gives p≈0.06 (one-sided), and roughly 19% of individual runs would fall below 22.8 under a normal approximation. The 'rank 14th' claim is therefore not statistically robust; it may reflect a favorable draw. Please report per-run scores/ranks and a significance test, or provide substantially more runs.
  2. [§3.1, footnote on p.1] The scores are assigned by the authors' human experts, one of whom is a co-author (footnote: IPhO 2021 gold medalist, IPhO 2022 marker). No raw agent outputs, per-item grading sheets, or independent grade audits are released. Since the central claim is a comparison to human gold medalists, even unintentional grading leniency could flip the conclusion. Please provide either independent, blinded grading or release the full answer transcripts and marking sheets for each run so the scoring can be audited.
  3. [Table 1 vs. Table 2] The 'top 10% on all three problems' claim is also fragile. For Theory 1, the mean is 9.02 ± 0.11 against an approximate top-10% threshold of ~9.0; the 0.02-point margin is far below the standard error. Theory 3 shows a 0.4-point margin (8.40 ± 0.27 vs. ~8.0), which is within noise. Only Theory 2 has a clear margin (6.08 ± 0.77 vs. ~5.0). Please provide the exact top-10% thresholds and the per-run distribution of per-problem scores.
  4. [Table 3] The ablation results are not statistically supported. For example, removing the AnswerReviewer gives 22.4 ± 0.6 vs. 23.5 ± 0.8; removing the ImageAnalyzer gives 22.8 ± 0.6. No significance tests are reported, and some differences are within the combined noise. Since the paper claims that 'principled tool integration' is the source of the improvement, these comparisons need additional runs or significance tests.
minor comments (4)
  1. [Table 2 caption, §3.2] Typographical errors: 'mediam' in the Table 2 caption and 'medium theory score' in §3.2 and §2.2 should be 'median'. The intended meaning is clear but these are distracting.
  2. [Table 1] The 'Human Top 10%' scores are presented as approximate (e.g., ~9.0, ~5.0, ~8.0). Please give the exact official thresholds or state explicitly that they are approximate.
  3. [§4.3] The WolframAlpha experiments use 10 self-generated problems, not IPhO problems. This is fine as an auxiliary study, but the paper should more clearly state that these results do not inform the main IPhO claims.
  4. [References] The reference list is extremely long; the full author list of the Gemini 2.5 technical report occupies several pages. Consider citing the model card more concisely or moving the full author list to an online supplement.

Circularity Check

0 steps flagged

No circular derivation: the benchmark target (IPhO 2025 problems, official marking scheme, human score distribution) is external, and the agent score is measured rather than fitted.

full rationale

The paper's central claim is an evaluation result against an independent, externally published benchmark: the IPhO 2025 theory problems, official scoring criteria obtained from https://ipho.olimpicos.net/, and official human contestant scores. The agent's 23.5/30 mean score is produced by running Physics Supernova on those problems and grading its outputs according to the external marking scheme; the score is not defined in terms of any quantity fitted to the benchmark, and no equation in the paper reduces the reported rank to a fitted parameter. The only self-referential elements are (i) the design-philosophy citation to the authors' prior Alita work [21] in Section 2.1, which is not load-bearing for the benchmark result, and (ii) the fact that human experts who grade the agent outputs include previous Olympiad medalists (Section 3.1 and title footnote). The latter is a genuine measurement-transparency/conflict-of-interest concern, but it is not a circularity: grading under an external rubric is a measurement, not a construction that makes the predicted quantity equal to its input. Likewise, the 0.7-point margin over the gold-medal median being smaller than the 0.8-point run-to-run SD is a statistical robustness issue, not a circularity. Because the benchmark and scoring rubric are external and the central claim does not reduce by construction to a fit or to a self-citation, no circular step is exhibited; score 0.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The paper's central numbers rely on assumptions about model knowledge timing, the integrity and correctness of the official marking scheme as applied by the authors' graders, and the effectiveness of the two prompt-based tools. There are no numerical free parameters fitted to data in the physics sense; the reported scores are measurements, not fitted values. No new physical entities are introduced.

axioms (3)
  • domain assumption Gemini 2.5 Pro's internal physics knowledge was acquired before IPhO 2025 and does not include the official solutions to these specific problems (no contamination).
    The benchmark's validity as a measure of reasoning, not memorization, depends on this; the paper asserts the problems are 'newly released' but does not test for memorization (Section 1, Section 3.1).
  • domain assumption The official IPhO 2025 marking schemes from ipho.olimpicos.net are the correct and complete scoring standard, and the authors' human experts apply them accurately and without bias.
    All scores and ablations derive from this human grading step; no inter-rater checks or grade sheets are provided (Section 3.1, Section A).
  • ad hoc to paper The ImageAnalyzer and AnswerReviewer tools, both implemented as Gemini 2.5 Pro with task-specific prompts, provide sufficiently accurate readings and reviews to support the observed improvements.
    The paper describes them as 'dedicated' tools but does not name a separate model for ImageAnalyzer; reproducibility depends on the exact prompts and model configuration (Section 2.2, Appendix B).

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics Supernova: AI Agent Matches Elite Gold Medalists at IPhO 2025." pith.science (2026). https://pith.science/paper/4DSWE7FX

@misc{pith2026250901659,
  author       = {Pith},
  title        = {Pith review of: Physics Supernova: AI Agent Matches Elite Gold Medalists at IPhO 2025},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4DSWE7FX}},
  note         = {Machine review of arXiv:2509.01659}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Physics provides fundamental laws that describe and predict the natural world. AI systems aspiring toward more general, real-world intelligence must therefore demonstrate strong physics problem-solving abilities: to formulate and apply physical laws for explaining and predicting physical processes. The International Physics Olympiad (IPhO)--the world's most prestigious physics competition--offers a rigorous benchmark for this purpose. We introduce Physics Supernova, an AI agent system with superior physics problem-solving abilities that match elite IPhO gold medalists. In IPhO 2025 theory problems, Physics Supernova attains 23.5/30 points, ranking 14th of 406 contestants and surpassing the median performance of human gold medalists. We extensively analyzed Physics Supernova's capabilities and flexibility across diverse physics tasks. These results show that principled tool integration within agent systems can deliver competitive improvements in solving challenging science problems. The codes are available at https://github.com/CharlesQ9/Physics-Supernova.

Figures

Figures reproduced from arXiv: 2509.01659 by Hongru Wang, Jiahao Qiu, Jiayi Geng, Jingzhe Shi, Mengdi Wang, Sanfeng Wu, Shilong Liu, Xinzhe Juan, Zelin Zhao.

Figure 1
Figure 1. Figure 1: Our proposed agent system: Physics Supernova, for solving theory problems in physics. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effect of ImageAnalyzer on Theory Problem 1 Part C. The problem here requires accurate measurement from a figure (shown left). We show 5 repeated experiments of directly applying LLMs and using Image Analyzer Tool (shown right), with reading difference > 0.01MHZ colored in red. As shown, the improvement mainly comes from a reduction in measurement error. 4.2 Case Study: Image Analyzer Tool We zoom in on Th… view at source ↗
Figure 1.1
Figure 1.1. Figure 1.1: Electric circuit for the experiment. Caution: one of the sample capacitors contains a dielectric with dielectric permittivity that depends on the capacitor voltage change rate. To keep this rate as stable as possible, when measuring at the positive voltages, the capacitor should be charged from 9 V down to –9 V, while measurements at the negative voltages should be done when capacitor is charged from –9 … view at source ↗
Figure 5
Figure 5. Figure 5: Left: an instrument-based experiment example (IPhO 2021 experiment problem 1); Right: a program-based [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 6 Pith papers

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

  1. Evaluating Large Language Models in Scientific Discovery

    cs.AI 2025-12 unverdicted novelty 8.0

    The SDE benchmark shows LLMs lag on scientific discovery tasks relative to general science tests, with diminishing scaling returns and shared weaknesses across models.

  2. Deep Research in Physical Sciences: A Multi-Agent Framework and Comprehensive Benchmark

    physics.comp-ph 2026-06 unverdicted novelty 7.0

    PhySciBench benchmark shows current AI models achieve at most 33.5% accuracy on physical science tasks; DelveAgent framework improves accuracy by up to 7.5 points and cuts costs to one-third.

  3. Can AI Evaluate AI Scientists? A Benchmarking Study of Autonomous Research Generation Systems Using Automated Multi-Model Review

    cs.AI 2026-04 reject novelty 6.0

    In a benchmark of four AI Scientist systems on 15 FARS proposals, LLM reviewers rated FARS's own papers about twice as high as Sakana v1/v2, CycleResearcher, and Data-to-Paper outputs, but the evaluation lacks human v...

  4. Towards Verifiable and Self-Correcting AI Physicists for Quantum Many-Body Simulations

    physics.comp-ph 2026-03 unverdicted novelty 6.0

    QMP-Bench supplies a realistic test set for AI on quantum many-body problems while PhysVEC uses integrated verifiers to turn unreliable LLM generations into code that passes both syntax and physics checks, outperformi...

  5. Students' Epistemological Beliefs and their Chatbot Preferences in AI-mediated Physics Learning

    physics.ed-ph 2026-07 conditional novelty 5.0

    Students who chose a combination chatbot (guided inquiry then answers) scored slightly higher on an epistemological-beliefs survey than answer-preference students, but the difference was not robust to multiple-testing...

  6. Aligning Perception, Reasoning, Modeling and Interaction: A Survey on Physical AI

    cs.AI 2025-10 unverdicted novelty 4.0

    A survey of physical AI that distinguishes theoretical physics reasoning from applied understanding and synthesizes advances in symbolic reasoning, embodied systems, and generative models to advocate for physics-groun...

Reference graph

Works this paper leans on

62 extracted references · 53 canonical work pages · cited by 6 Pith papers

  1. [1]

    Richard P. Feynman. The Character of Physical Law. MIT Press, 1965

    work page 1965

  2. [2]

    P. J. E. Peebles. Principles of Physical Cosmology. Princeton University Press, 1993

    work page 1993

  3. [3]

    Eugene P. Wigner. The unreasonable effectiveness of mathematics in the natural sciences. Communications on Pure and Applied Mathematics, 13(1):1–14, 1960

    work page 1960

  4. [4]

    Invariante variationsprobleme

    Emmy Noether. Invariante variationsprobleme. Nachrichten von der Gesellschaft der Wissenschaften zu Göttingen, Mathematisch-Physikalische Klasse, pages 235–257, 1918

    work page 1918

  5. [5]

    Åström and Richard M

    Karl J. Åström and Richard M. Murray.Feedback Systems: An Introduction for Scientists and Engineers. Princeton University Press, 2008

    work page 2008

  6. [6]

    Edward A. Lee. Cyber physical systems: Design challenges. In Proc. IEEE Int’l Symposium on Object/Component/Service-Oriented Real-Time Distributed Computing (ISORC), pages 363–369, 2008

    work page 2008

  7. [7]

    Superintelligence: Paths, Dangers, Strategies

    Nick Bostrom. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014

    work page 2014

  8. [8]

    A path towards autonomous machine intelligence

    Yann LeCun. A path towards autonomous machine intelligence. OpenReview position paper, 2022

    work page 2022

  9. [9]

    Official Website: https://www.ipho-new.org/, ongoing

    International Physics Olympiad (IPhO). Official Website: https://www.ipho-new.org/, ongoing. Accessed: 2025-07-29

    work page 2025

  10. [10]

    Official Website: https://www.imo-official.org/, ongoing

    International Mathematical Olympiad (IMO). Official Website: https://www.imo-official.org/, ongoing. Accessed: 2025-07-29

    work page 2025

  11. [11]

    Official Website: https://www.ipho2025.fr/, ongoing

    International Physics Olympiad (IPhO) 2025. Official Website: https://www.ipho2025.fr/, ongoing. Ac- cessed: 2025-07-29

    work page 2025

  12. [12]

    Seephys: Does seeing help thinking? – benchmarking vision-based physics reasoning, 2025

    Kun Xiang, Heng Li, Terry Jingchen Zhang, Yinya Huang, Zirong Liu, Peixin Qu, Jixi He, Jiaqi Chen, Yu-Jie Yuan, Jianhua Han, Hang Xu, Hanhui Li, Mrinmaya Sachan, and Xiaodan Liang. Seephys: Does seeing help thinking? – benchmarking vision-based physics reasoning, 2025

    work page 2025

  13. [13]

    Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems, 2024

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, Jie Liu, Lei Qi, Zhiyuan Liu, and Maosong Sun. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems, 2024

    2024

  14. [14]

    Phybench: Holistic evaluation of physical perception and reasoning in large language models, 2025

    Shi Qiu, Shaoyang Guo, Zhuo-Yang Song, Yunbo Sun, Zeyu Cai, Jiashen Wei, Tianyu Luo, Yixuan Yin, Haoxu Zhang, Yi Hu, Chenyang Wang, Chencheng Tang, Haoling Chang, Qi Liu, Ziheng Zhou, Tianyu Zhang, Jingtian Zhang, Zhangyi Liu, Minghao Li, Yuku Zhang, Boxuan Jing, Xianqi Yin, Yutong Ren, Zizhuo Fu, Jiaming Ji, Weike Wang, Xudong Tian, Anqi Lv, Laifu Man, J...

    work page 2025

  15. [15]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effective than scaling model parameters, 2024

    2024

  16. [16]

    Scaling inference computation: Compute-optimal inference for problem-solving with language models

    Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. Scaling inference computation: Compute-optimal inference for problem-solving with language models. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS’24, 2024

    work page 2024

  17. [17]

    A survey of the evolution of language model-based dialogue systems: Data, task and models, 2025

    Hongru Wang, Lingzhi Wang, Yiming Du, Liang Chen, Jingyan Zhou, Yufei Wang, and Kam-Fai Wong. A survey of the evolution of language model-based dialogue systems: Data, task and models, 2025

    work page 2025

  18. [18]

    Large language models are zero-shot reasoners, 2023

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners, 2023

    2023

  19. [19]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, 2023

    2023

  20. [20]

    A survey of self-evolving agents: On path to artificial super intelligence, 2025

    Huan ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, Qihan Ren, Cheng Qian, Zhenhailong Wang, Minda Hu, Huazheng Wang, Qingyun Wu, Heng Ji, and Mengdi Wang. A survey of se...

    work page 2025

  21. [21]

    Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution, 2025

    Jiahao Qiu, Xuan Qi, Tongcheng Zhang, Xinzhe Juan, Jiacheng Guo, Yifu Lu, Yimin Wang, Zixin Yao, Qihan Ren, Xun Jiang, Xing Zhou, Dongrui Liu, Ling Yang, Yue Wu, Kaixuan Huang, Shilong Liu, Hongru Wang, and Mengdi Wang. Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution, 2025

    2025

  22. [22]

    Group-in-group policy optimization for llm agent training, 2025

    Lang Feng, Zhenghai Xue, Tingcong Liu, and Bo An. Group-in-group policy optimization for llm agent training, 2025

    2025

  23. [23]

    Agentdistill: Training-free agent distillation with generalizable mcp boxes, 2025

    Jiahao Qiu, Xinzhe Juan, Yimin Wang, Ling Yang, Xuan Qi, Tongcheng Zhang, Jiacheng Guo, Yifu Lu, Zixin Yao, Hongru Wang, Shilong Liu, Xun Jiang, Liu Leqi, and Mengdi Wang. Agentdistill: Training-free agent distillation with generalizable mcp boxes, 2025

    work page 2025

  24. [24]

    Gaia: a benchmark for general ai assistants

    Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants. In The Twelfth International Conference on Learning Representations, 2023

    2023

  25. [25]

    Advanced version of gemini with deep think officially achieves gold-medal standard at the international mathematical olympiad

    Thang Luong and Edward Lockhart. Advanced version of gemini with deep think officially achieves gold-medal standard at the international mathematical olympiad. Google DeepMind Blog, July 2025

    work page 2025

  26. [26]

    On path to multimodal historical reasoning: Histbench and histagent, 2025

    Jiahao Qiu, Fulian Xiao, Yimin Wang, Yuchen Mao, Yijia Chen, Xinzhe Juan, Shu Zhang, Siran Wang, Xuan Qi, Tongcheng Zhang, Zixin Yao, Jiacheng Guo, Yifu Lu, Charles Argon, Jundi Cui, Daixin Chen, Junran Zhou, Shuyao Zhou, Zhanpeng Zhou, Ling Yang, Shilong Liu, Hongru Wang, Kaixuan Huang, Xun Jiang, Yuming Cao, Yue Chen, Yunfei Chen, Zhengyi Chen, Ruowei D...

    work page 2025

  27. [27]

    Wolfram|alpha

    Wolfram Alpha LLC. Wolfram|alpha. https://www.wolframalpha.com/, 2009. Accessed: 2025-08-12

    work page 2009

  28. [28]

    ‘smo- lagents‘: a smol library to build great agentic systems

    Aymeric Roucher, Albert Villanova del Moral, Thomas Wolf, Leandro von Werra, and Erik Kaunismäki. ‘smo- lagents‘: a smol library to build great agentic systems. https://github.com/huggingface/smolagents, 2025

    work page 2025

  29. [29]

    Yichen Huang and Lin F. Yang. Gemini 2.5 pro capable of winning gold at imo 2025, 2025

    work page 2025

  30. [30]

    Lean-star: Learning to interleave thinking and proving, 2025

    Haohan Lin, Zhiqing Sun, Sean Welleck, and Yiming Yang. Lean-star: Learning to interleave thinking and proving, 2025

    work page 2025

  31. [31]

    ultraviolet catastrophe

    Lord Rayleigh and J. H. Jeans. On the theory of quantized matter and radiation – the rayleigh–jeans law and the ultraviolet catastrophe. Philosophical Magazine, 1900–1905. Classical law predicting divergent energy at short wavelengths (“ultraviolet catastrophe”)

    work page 1900

  32. [32]

    On the law of distribution of energy in the normal spectrum

    Max Planck. On the law of distribution of energy in the normal spectrum. Annalen der Physik, 4:553–563, 1900. Introduction of quantized energy elements, resolving the ultraviolet catastrophe

    work page 1900

  33. [33]

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, Luke Marris, Sam Petulla, Colin Gaffney, Asaf Aharoni, Nathan Lintz, Tiago Cardal Pais, Henrik Jacobsson, Idan Szpektor, Nan-Jiang Jiang, Krishna Haridasan, Ahmed Om- ran, Nikunj Saunshi, Dara Bahri, Gaurav Mi...

    work page 2025

  34. [34]

    Language models are unsupervised multitask learners

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019

    2019

  35. [35]

    Toward a theory of agents as tool-use decision-makers, 2025

    Hongru Wang, Cheng Qian, Manling Li, Jiahao Qiu, Boyang Xue, Mengdi Wang, Heng Ji, and Kam-Fai Wong. Toward a theory of agents as tool-use decision-makers, 2025

    work page 2025

  36. [36]

    Autogen: Enabling next-gen llm applications via multi-agent conversation, 2023

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation, 2023

    2023

  37. [37]

    Introducing deep research

    OpenAI. Introducing deep research

    work page

  38. [38]

    Manus, 2024

    Manus Team. Manus, 2024

    work page 2024

  39. [39]

    Acting less is reasoning more! teaching model to act efficiently, 2025

    Hongru Wang, Cheng Qian, Wanjun Zhong, Xiusi Chen, Jiahao Qiu, Shijue Huang, Bowen Jin, Mengdi Wang, Kam-Fai Wong, and Heng Ji. Acting less is reasoning more! teaching model to act efficiently, 2025

    work page 2025

  40. [40]

    Agentic reinforced policy optimization, 2025

    Guanting Dong, Hangyu Mao, Kai Ma, Licheng Bao, Yifei Chen, Zhongyuan Wang, Zhongxia Chen, Jiazhen Du, Huiyang Wang, Fuzheng Zhang, Guorui Zhou, Yutao Zhu, Ji-Rong Wen, and Zhicheng Dou. Agentic reinforced policy optimization, 2025

    work page 2025

  41. [41]

    Qiu, and Yuqing Yang

    Xufang Luo, Yuge Zhang, Zhiyuan He, Zilong Wang, Siyun Zhao, Dongsheng Li, Luna K. Qiu, and Yuqing Yang. Agent lightning: Train any ai agents with reinforcement learning, 2025

    work page 2025

  42. [42]

    Ai achieves silver-medal standard solving international mathematical olympiad problems

    AlphaProof and AlphaGeometry teams. Ai achieves silver-medal standard solving international mathematical olympiad problems. Google DeepMind Blog, July 2024. Published July 25, 2024

    work page 2024

  43. [43]

    Emoagent: Assessing and safeguarding human-ai interaction for mental health safety, 2025

    Jiahao Qiu, Yinghui He, Xinzhe Juan, Yimin Wang, Yuhan Liu, Zixin Yao, Yue Wu, Xun Jiang, Ling Yang, and Mengdi Wang. Emoagent: Assessing and safeguarding human-ai interaction for mental health safety, 2025

    work page 2025

  44. [44]

    Seed-prover: Deep and broad reasoning for automated theorem proving, 2025

    Luoxin Chen, Jinming Gu, Liankai Huang, Wenhao Huang, Zhicheng Jiang, Allan Jie, Xiaoran Jin, Xing Jin, Chenggang Li, Kaijing Ma, Cheng Ren, Jiawei Shen, Wenlei Shi, Tong Sun, He Sun, Jiahui Wang, Siran Wang, Zhihong Wang, Chenrui Wei, Shufa Wei, Yonghui Wu, Yuchen Wu, Yihang Xia, Huajian Xin, Fan Yang, Huaiyuan Ying, Hongyi Yuan, Zheng Yuan, Tianyang Zha...

    work page 2025

  45. [45]

    Physicsarena: The first multimodal physics reasoning benchmark exploring variable, process, and solution dimensions, 2025

    Song Dai, Yibo Yan, Jiamin Su, Dongfang Zihao, Yubo Gao, Yonghua Hei, Jungang Li, Junyan Zhang, Sicheng Tao, Zhuoran Gao, and Xuming Hu. Physicsarena: The first multimodal physics reasoning benchmark exploring variable, process, and solution dimensions, 2025

    work page 2025

  46. [46]

    Physreason: A comprehensive benchmark towards physics-based reasoning, 2025

    Xinyu Zhang, Yuxuan Dong, Yanrui Wu, Jiaxing Huang, Chengyou Jia, Basura Fernando, Mike Zheng Shou, Lingling Zhang, and Jun Liu. Physreason: A comprehensive benchmark towards physics-based reasoning, 2025

    work page 2025

  47. [47]

    Philosophiæ Naturalis Principia Mathematica

    Isaac Newton. Philosophiæ Naturalis Principia Mathematica. Jussu Societatis Regiæ ac Typis Josephi Streater, 1687

    work page

  48. [48]

    Salib, Luke Huang, Dirk R

    Shiekh Zia Uddin, Sachin Vaidya, Shrish Choudhary, Zhuo Chen, Raafat K. Salib, Luke Huang, Dirk R. Englund, and Marin Soljaˇci´c. Ai-driven robotics for free-space optics, 2025

    work page 2025

  49. [49]

    Lean4lean: Towards a verified typechecker for lean, in lean, 2024

    Mario Carneiro. Lean4lean: Towards a verified typechecker for lean, in lean, 2024

    work page 2024

  50. [50]

    Lean copilot: Large language models as copilots for theorem proving in lean, 2025

    Peiyang Song, Kaiyu Yang, and Anima Anandkumar. Lean copilot: Large language models as copilots for theorem proving in lean, 2025

    work page 2025

  51. [51]

    Bobbin, Samiha Sharlin, Parivash Feyzishendi, An Hong Dang, Catherine M

    Maxwell P. Bobbin, Samiha Sharlin, Parivash Feyzishendi, An Hong Dang, Catherine M. Wraback, and Tyler R. Josephson. Formalizing chemical physics using the lean theorem prover, 2023

    work page 2023

  52. [52]

    Smolensky

    P. Smolensky. Connectionist ai, symbolic ai, and the brain. Artificial Intelligence Review, 1(3):95–109, 1987

    work page 1987

  53. [53]

    The lean 4 theorem prover and programming language (system description)

    Leonardo de Moura and Sebastian Ullrich. The lean 4 theorem prover and programming language (system description). In Automated Deduction – CADE 28, pages 625–635. Springer, Cham, 2021

    work page 2021

  54. [54]

    Z. Z. Ren, Zhihong Shao, Junxiao Song, Huajian Xin, Haocheng Wang, Wanjia Zhao, Liyue Zhang, Zhe Fu, Qihao Zhu, Dejian Yang, Z. F. Wu, Zhibin Gou, Shirong Ma, Hongxuan Tang, Yuxuan Liu, Wenjun Gao, Daya Guo, and Chong Ruan. Deepseek-prover-v2: Advancing formal mathematical reasoning via reinforcement learning for subgoal decomposition, 2025

    work page 2025

  55. [55]

    Kimina-prover preview: Towards large formal reasoning models with reinforcement learning, 2025

    Numina & Kimi Team. Kimina-prover preview: Towards large formal reasoning models with reinforcement learning, 2025

    work page 2025

  56. [56]

    Solving formal math problems by decomposition and iterative reflection, 2025

    Yichi Zhou, Jianqiu Zhao, Yongxin Zhang, Bohan Wang, Siran Wang, Luoxin Chen, Jiahui Wang, Haowei Chen, Allan Jie, Xinbo Zhang, Haocheng Wang, Trung Luong, Rong Ye, Phan Nhat Hoang, Huishuai Zhang, Peng Sun, and Hang Li. Solving formal math problems by decomposition and iterative reflection, 2025

    work page 2025

  57. [57]

    Prover agent: An agent-based framework for formal mathematical proofs, 2025

    Kaito Baba, Chaoran Liu, Shuhei Kurita, and Akiyoshi Sannai. Prover agent: An agent-based framework for formal mathematical proofs, 2025

    work page 2025

  58. [58]

    Apollo: Automated llm and lean collaboration for advanced formal reasoning, 2025

    Azim Ospanov, Farzan Farnia, and Roozbeh Yousefzadeh. Apollo: Automated llm and lean collaboration for advanced formal reasoning, 2025. 20 A Examples of IPhO 2025 Problems Scoring Criteria We provide two examples of the IPhO 2025 scoring criteria (corresponding to Theory Problem 1 Part C.1 and Theory Problem 3 Part C.2, respectively) in Figure 4, obtained...

    work page 2025

  59. [59]

    At 𝑇0 = 6 °C a grey-white fog appears while opening the bottle

    work page

  60. [60]

    At 𝑇0 = 6 °C a blue fog appears while opening the bottle

    work page

  61. [61]

    At 𝑇0 = 20 °C a grey-white fog appears while opening the bottle

    work page

  62. [62]

    You are an expert in dealing with image in Physics O lym pi ad s

    At 𝑇0 = 20 °C a blue fog appears while opening the bottle. 0.7pt SOLUTION: C.2.1. The adiabatic reversible expansion goes from 𝑃𝑖 to 𝑃0. C.2.2. 𝑇𝑓 = 𝑇0 ⒧ 𝑃𝑖 𝑃0 ⒭ (1/𝛾)−1 C.2.3. For 𝑇0 = 6 °C: 𝑃𝑖 = 4.69 bar and 𝑇𝑓 = 195.3 K = −77.8 °C. C.2.4. For 𝑇0 = 20 °C: 𝑃𝑖 = 7.45 bar and 𝑇𝑓 = 184.3 K = −88.8 °C. C.2.5. First method: comparison 𝑃sat(𝑇𝑓) and 𝑃𝑓 = 𝑃0. Se...

    work page 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.