Pith. sign in

REVIEW 3 major objections 3 minor 1 cited by

Mimicking the Physicist's Eye:A VLM-centric Approach for Physics Formula Discovery

T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read VIPER-R1 discovers symbolic physics formulas by combining visual perception with symbolic reasoning.

desk verdict Plausible new integration of vision, trajectory, and symbolic regression for formula discovery, but the abstract's headline comparison is confounded by the external SR tool. read the letter →

arxiv 2508.17380 v1 pith:XDH3YYOX submitted 2025-08-24 cs.AI

Jiaqi Liu , Songning Lai , Pengze Li , Di Yu , Wenjie Zhou , Yiyang Zhou , Peng Xia , Zijun Wang
show 7 more authors
Xi Chen Shixiang Tang Lei Bai Wanli Ouyang Mingyu Ding Huaxiu Yao Aoran Wang
This is my paper · ORCID
classification cs.AI
keywords physicsformuladiscoveryvision-languagemodelssymbolicregressionphaseportraitsreinforcementlearningmultimodalcorpusscientificcausalchainofthought
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

The paper sets out to show that automated physics-formula discovery is substantially better when the system can see motion, not just read numbers. It presents VIPER-R1, a vision-language model that looks at kinematic phase portraits, states a symbolic hypothesis through a causal chain of thought, then refines the formula against trajectory data with an external symbolic-regression tool. The authors introduce PhysSymbol, a 5,000-instance multimodal corpus, and report that VIPER-R1 consistently outperforms state-of-the-art VLM baselines in both accuracy and interpretability. If the claim holds, formula discovery moves from uni-modal symbolic regression and text-only LLM reasoning toward a perception-and-reasoning loop closer to a physicist's own practice.

What carries the argument

The load-bearing mechanism is the three-stage training-and-inference curriculum. Motion Structure Induction (MSI) teaches the model to interpret kinematic phase portraits; the Causal Chain of Thought (C-CoT) makes the model articulate a step-by-step physical story before writing a formula; Reward-Guided Symbolic Calibration (RGSC) treats formula structure as a policy to be refined by reinforcement learning; and Symbolic Residual Realignment ($SR^{2}$), the inference-time step, calls an external symbolic regression tool to adjust the ansatz against numeric residuals. The division of labor matters: the VLM supplies the symbolic form and causal interpretation, while the external tool supplies numerical precision.

What would settle it

Take the trained VIPER-R1 and feed it a real video of a simple pendulum or projectile motion with tracked trajectories; if the recovered formula and fitted constants deviate materially from the known physical law while matching the corpus's synthetic distribution in style, the claim that the model discovers physical laws from observational data would be falsified.

Watch

Extended reading notes

Core claim

VIPER-R1 is a multimodal model that performs Visual Induction for Physics-based Equation Reasoning. It claims that the missing ingredient in current formula-discovery systems is visual perception: the rich spatio-temporal structure of motion, encoded in phase portraits, carries information that uni-modal data omits. Trained with Motion Structure Induction, the model learns to read those portraits, to propose hypotheses through a Causal Chain of Thought, and to calibrate its symbolic guesses with reward-guided reinforcement learning. At inference it acts as an agent, first committing to a high-confidence symbolic ansatz, then invoking an external symbolic regression tool for Symbolic Residual Realignment, a step the authors liken to perturbative correction in physics. The paper's central assertion is that this pipeline beats state-of-the-art VLM baselines on accuracy and interpretability, enabling more precise discovery of physical laws.

Load-bearing premise

The load-bearing premise is that the 5,000-instance PhysSymbol corpus is representative enough of real-world motion that the accuracy reported on it carries over to observational data outside the corpus.

Editorial extensions

If this is right

  • Accuracy of recovered formulas improves because visual phase portraits expose spatio-temporal patterns unavailable to numeric-only or text-only inputs.
  • Scientists can audit a discovery: C-CoT yields a readable causal chain from observed motion to hypothesized law, and SR^2 shows where the empirical data still deviate from the symbolic form.
  • The trained model can approximate a physicist's perturbation analysis: it starts from a global ansatz and corrects local residuals instead of searching the whole formula space from scratch.
  • PhysSymbol becomes a reusable benchmark for multimodal physical-law discovery, allowing later systems to be trained and compared on the same visual-plus-trajectory task.

Reading between the lines

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

  • Beyond the corpus, a natural test is to run VIPER-R1 on real video of a pendulum, falling body, or spring and compare the recovered equation and constants to the known law; success would confirm that the visual channel transfers beyond simulated phase portraits.
  • The architecture suggests a general recipe: let a language model propose symbolic structure and a numeric tool fit the coefficients. Iterating this loop on higher-dimensional or coupled systems could extend the method beyond the kinematics examples in the corpus.
  • Because all training and evaluation rest on PhysSymbol, the method's ceiling is set by the corpus's diversity; broadening it to fluids, waves, or electromagnetic data would show whether the mechanism is a general discovery engine or a kinematic special case.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The manuscript proposes VIPER-R1, a multimodal vision-language model for automated physics formula discovery from visual observations of motion. The proposed pipeline trains the model with a curriculum of Motion Structure Induction, supervised fine-tuning with a Causal Chain of Thought, and reinforcement-learning-based Reward-Guided Symbolic Calibration; at inference the model first proposes a symbolic ansatz and then invokes an external symbolic regression tool for Symbolic Residual Realignment. The paper also introduces PhysSymbol, a 5,000-instance multimodal corpus, and claims that VIPER-R1 consistently outperforms state-of-the-art VLM baselines in accuracy and interpretability. The manuscript as submitted consists only of the abstract and a project-page link; no methods, experimental details, or quantitative results are provided.

Significance. If the empirical claims were substantiated, the work would be a meaningful step toward integrating visual perception into symbolic regression, a direction that is currently underrepresented; the proposed two-stage ansatz-plus-residual-alignment protocol is a sensible design that could be tested with controlled ablations. The introduction of PhysSymbol would also be a useful community resource. However, the manuscript as submitted contains no verifiable evidence: the central claims are stated without any metrics, baselines, error bars, ablation studies, or external validation, so the significance cannot currently be assessed.

major comments (3)
  1. [Abstract (central claim)] The headline claim that VIPER-R1 'consistently outperforms state-of-the-art VLM baselines in accuracy and interpretability' is stated without any quantitative support in the manuscript; no accuracy metric, no baseline names, no error bars, no dataset splits, and no ablations appear anywhere. This is load-bearing because the paper's contribution is empirical, and the claim cannot be verified or reproduced from the text.
  2. [Abstract (inference protocol and SR^2)] The described inference protocol gives VIPER-R1 access to an external symbolic regression tool (SR^2) after it posits an ansatz, while no baseline is described as having an equivalent tool. Since SR^2 could, by itself, fit many formulas from residual trajectory data, the reported advantage may be entirely due to tool augmentation rather than to the visual induction, C-CoT, or RGSC components. The authors must either grant the same tool to all baselines or ablate SR^2 from VIPER-R1 to isolate the model's contribution.
  3. [Abstract (PhysSymbol corpus)] The PhysSymbol corpus is introduced as the sole evaluation basis, but the abstract gives no information about how the 5,000 instances were generated, what physical phenomena they cover, what the difficulty distribution is, or how the corpus is split. The abstract also provides no external validation on real experimental data, so the claim that VIPER-R1 enables 'more precise discovery of physical laws' is unsupported beyond a self-contained synthetic benchmark.
minor comments (3)
  1. [Abstract (terminology)] The phrase 'consistently outperforms' implies repeated trials or multiple benchmarks, but no evidence of variance or multiple settings is provided; please define 'consistent' operationally, for example with standard errors or multiple runs.
  2. [Abstract (interpretability)] The term 'interpretability' is used without an operational definition; please specify whether it is measured by human judgment, faithfulness of the Chain of Thought, formula simplicity, or another criterion, and state the corresponding metric.
  3. [General] The manuscript appears to be an abstract only; the full paper with the complete methodology, training details, baseline comparisons, and results is required for a substantive review.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity is established from the abstract: the SR² step uses data but is not shown to be a disguised prediction.

full rationale

The provided material contains no equations, no derived chain, and no claim that a quantity is predicted from a first-principles result. The only step that could resemble circularity is the inference-time Symbolic Residual Realignment (SR²), in which the model posits an ansatz and an external symbolic regression tool reconciles the theoretical model with empirical data. That is a fitting step, but the abstract presents it as part of the discovery pipeline, not as an independent prediction of the data it fits. No evaluation protocol is given, so we cannot exhibit a specific reduction showing that the reported accuracy is statistically forced by a fitted parameter. There are no self-citations, no uniqueness theorems imported from the authors, and no ansatz smuggled in via citation. The concern that VLM baselines may lack the SR² tool is an attribution/ablation concern about the comparison, not a circularity of the derivation. Therefore the honest finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

No explicit free parameters or invented physical entities are mentioned in the abstract. The two domain assumptions listed are necessary for the claimed transfer from the training corpus to real-world physics discovery.

assumptions (2)
  • domain assumption Kinematic phase portraits contain enough information to infer the underlying symbolic law.
    The MSI curriculum trains the model to map phase portraits to equation structures; this mapping is assumed learnable and is not proven in the abstract.
  • domain assumption The PhysSymbol corpus is representative of real-world dynamic phenomena.
    All reported gains are measured on this 5,000-instance corpus, yet the abstract offers no external validation on real experimental data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mimicking the Physicist's Eye:A VLM-centric Approach for Physics Formula Discovery." pith.science (2026). https://pith.science/paper/XDH3YYOX

@misc{pith2026250817380,
  author       = {Pith},
  title        = {Pith review of: Mimicking the Physicist's Eye:A VLM-centric Approach for Physics Formula Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XDH3YYOX}},
  note         = {Machine review of arXiv:2508.17380}
}
read the original abstract

Automated discovery of physical laws from observational data in the real world is a grand challenge in AI. Current methods, relying on symbolic regression or LLMs, are limited to uni-modal data and overlook the rich, visual phenomenological representations of motion that are indispensable to physicists. This "sensory deprivation" severely weakens their ability to interpret the inherent spatio-temporal patterns within dynamic phenomena. To address this gap, we propose VIPER-R1, a multimodal model that performs Visual Induction for Physics-based Equation Reasoning to discover fundamental symbolic formulas. It integrates visual perception, trajectory data, and symbolic reasoning to emulate the scientific discovery process. The model is trained via a curriculum of Motion Structure Induction (MSI), using supervised fine-tuning to interpret kinematic phase portraits and to construct hypotheses guided by a Causal Chain of Thought (C-CoT), followed by Reward-Guided Symbolic Calibration (RGSC) to refine the formula structure with reinforcement learning. During inference, the trained VIPER-R1 acts as an agent: it first posits a high-confidence symbolic ansatz, then proactively invokes an external symbolic regression tool to perform Symbolic Residual Realignment (SR^2). This final step, analogous to a physicist's perturbation analysis, reconciles the theoretical model with empirical data. To support this research, we introduce PhysSymbol, a new 5,000-instance multimodal corpus. Experiments show that VIPER-R1 consistently outperforms state-of-the-art VLM baselines in accuracy and interpretability, enabling more precise discovery of physical laws. Project page: https://jiaaqiliu.github.io/VIPER-R1/

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. IRIS: A Real-World Benchmark for Inverse Recovery and Identification of Physical Dynamic Systems from Monocular Video

    cs.CV 2026-03 accept novelty 6.0 of 10

    IRIS releases 220 real 4K videos of eight dynamical systems with ground-truth parameters plus a protocol that measures parameter recovery, equation selection, and multi-body failure modes of unsupervised video-to-phys...

Reference graph

Works this paper leans on

60 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Introducing claude 4

    Anthropic. Introducing claude 4. https://www.anthropic.com/news/claude-4, 2024

    work page 2024

  3. [3]

    Claude 3.7 sonnet and claude code

    Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025

    2025

  4. [4]

    Qwen- VL : A versatile vision-language model for understanding, localization, text reading, and beyond

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen- VL : A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2309.16609, 2023

    arXiv 2023

  5. [6]

    Qwen2.5- VL technical report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5- VL technical report. arXiv preprint arXiv:2502.13923, 2025 b

    arXiv 2025

  6. [7]

    Neural symbolic regression that scales

    Luca Biggio, Tommaso Bendinelli, Alexander Neitz, Aurelien Lucchi, and Giambattista Parascandolo. Neural symbolic regression that scales. In International Conference on Machine Learning, pp.\ 936--945. Pmlr, 2021

    work page 2021

  7. [8]

    Evoprompting: Language models for code-level neural architecture search

    Angelica Chen, David Dohan, and David So. Evoprompting: Language models for code-level neural architecture search. volume 36, pp.\ 7787--7817, 2023

    work page 2023

  8. [9]

    Interpretable machine learning for science with pysr and symbolicregression.jl, May 2023

    Miles Cranmer. Interpretable machine learning for science with pysr and symbolicregression.jl, May 2023

    work page 2023

Show all 60 references
  1. [10]

    Gemini 2.5: Our most intelligent ai model

    Google. Gemini 2.5: Our most intelligent ai model. https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/#gemini-2-5-thinking, 2025

    2025

  2. [11]

    Symbolic regression with a learned concept library

    Arya Grayeli, Atharva Sehgal, Omar Costilla Reyes, Miles Cranmer, and Swarat Chaudhuri. Symbolic regression with a learned concept library. Advances in Neural Information Processing Systems, 37: 0 44678--44709, 2024

    2024

  3. [12]

    Connecting large language models with evolutionary algorithms yields powerful prompt optimizers

    Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yujiu Yang. Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532, 2023

    2023 arXiv

  4. [13]

    open-r1: Fully open reproduction of deepseek-r1

    Huggingface. open-r1: Fully open reproduction of deepseek-r1. https://github.com/huggingface/open-r1, 2025. [Online; accessed: 2025-01-24]

    2025

  5. [14]

    End-to-end symbolic regression with transformers

    Pierre-Alexandre Kamienny, St \'e phane d'Ascoli, Guillaume Lample, and Fran c ois Charton. End-to-end symbolic regression with transformers. volume 35, pp.\ 10269--10281, 2022

    2022

  6. [15]

    Genetic programming as a means for programming computers by natural selection

    John R Koza. Genetic programming as a means for programming computers by natural selection. Statistics and computing, 4: 0 87--112, 1994

    1994

  7. [16]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

    2023

  8. [17]

    Contemporary symbolic regression methods and their relative performance

    William La Cava, Patryk Orzechowski, Bogdan Burlacu, Fabr \' cio Olivetti de Fran c a, Marco Virgolin, Ying Jin, Michael Kommenda, and Jason H Moore. Contemporary symbolic regression methods and their relative performance. In Advances in Neural Information Processing Systems D...

    2021

  9. [18]

    Large language models as evolution strategies

    Robert Lange, Yingtao Tian, and Yujin Tang. Large language models as evolution strategies. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, pp.\ 579--582, 2024 a

    2024

  10. [19]

    Large language models as evolution strategies

    Robert Tjarko Lange, Yingtao Tang, and Yujin Tian. Large language models as evolution strategies. In Genetic and Evolutionary Computation Conference, pp.\ 1332--1340, 2024 b

    2024

  11. [20]

    Evolution through large models

    Joel Lehman, Jonathan Gordon, Shawn Jain, Kamal Ndousse, Cathy Yeh, and Kenneth O Stanley. Evolution through large models. In Handbook of evolutionary machine learning, pp.\ 331--366. Springer, 2023

    2023

  12. [21]

    Automated statistical model discovery with language models

    Michael Y Li, Emily B Fox, and Noah D Goodman. Automated statistical model discovery with language models. arXiv preprint arXiv:2402.17879, 2024

    2024 arXiv

  13. [22]

    Mllm-based discovery of intrinsic coordinates and governing equations from high-dimensional data, May 2025

    Ruikun Li, Yan Lu, Shixiang Tang, Biqing Qi, and Wanli Ouyang. Mllm-based discovery of intrinsic coordinates and governing equations from high-dimensional data, May 2025

    2025

  14. [23]

    The AI scientist: Towards fully automated open-ended scientific discovery

    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, 2024

    2024 arXiv

  15. [24]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35: 0 2507...

    2022

  16. [25]

    Tenenbaum, Daniela Rus, Chuang Gan, and Wojciech Matusik

    Pingchuan Ma, Tsun-Hsuan Wang, Minghao Guo, Zhiqing Sun, Joshua B. Tenenbaum, Daniela Rus, Chuang Gan, and Wojciech Matusik. LLM and simulation as bilevel optimizers: A new paradigm to advance physical scientific discovery. In Ruslan Salakhutdinov, Zico Kolter, Katherine Helle...

    2024

  17. [26]

    Llm and simulation as bilevel optimizers: A new paradigm to advance physical scientific discovery

    Pingchuan Ma, Tsun-Hsuan Wang, Minghao Guo, Zhiqing Sun, Joshua B Tenenbaum, Daniela Rus, Chuang Gan, and Wojciech Matusik. Llm and simulation as bilevel optimizers: A new paradigm to advance physical scientific discovery. arXiv preprint arXiv:2405.09783, 2024 b

    2024 arXiv

  18. [27]

    Data-driven discovery with large generative models

    Bodhisattwa Prasad Majumder, Harshit Surana, Dhruv Agarwal, Sanchaita Hazra, Ashish Sabharwal, and Peter Clark. Data-driven discovery with large generative models. arXiv preprint arXiv:2402.13610, 2024 a

    2024 arXiv

  19. [28]

    Discoverybench: Towards data-driven discovery with large language models

    Bodhisattwa Prasad Majumder, Harshit Surana, Dhruv Agarwal, Bhavana Dalvi Mishra, Abhijeetsingh Meena, Aryan Prakhar, Tirth Vora, Tushar Khot, Ashish Sabharwal, and Peter Clark. Discoverybench: Towards data-driven discovery with large language models. arXiv preprint arXiv:2407...

    2024 arXiv

  20. [29]

    Rethinking symbolic regression datasets and benchmarks for scientific discovery

    Yoshitomo Matsubara, Naoya Chiba, Ryo Igarashi, and Yoshitaka Ushiku. Rethinking symbolic regression datasets and benchmarks for scientific discovery. arXiv preprint arXiv:2206.10540, 2022

    2022 arXiv

  21. [30]

    SNIP : Bridging mathematical symbolic and numeric realms with unified pre-training

    Kazem Meidani, Parshin Agarwal, Mohammad Taha Bahadori, Jayant Liang, and Amir Barati Farimani. SNIP : Bridging mathematical symbolic and numeric realms with unified pre-training. In International Conference on Learning Representations, 2023

    2023

  22. [31]

    In-context symbolic regression: leveraging language models for function discovery

    Matteo Merler, Mirko Nanni, and Fabrizio Silvestri. In-context symbolic regression: leveraging language models for function discovery. arXiv preprint arXiv:2404.19094, 2024

    2024 arXiv

  23. [32]

    Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models

    Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar. Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229, 2024

    arXiv

  24. [33]

    Symbolic regression via deep reinforcement learning enhanced genetic programming seeding

    T Nathan Mundhenk, Mikel Landajuela, Ruben Glatt, Claudio P Santiago, Daniel M Faissol, and Brenden K Petersen. Symbolic regression via deep reinforcement learning enhanced genetic programming seeding. In Advances in Neural Information Processing Systems, pp.\ 24912--24923, 2021

    2021

  25. [34]

    Gpt-4o system card, 2024

    OpenAI. Gpt-4o system card, 2024. URL https://arxiv.org/abs/2410.21276

    2024 arXiv

  26. [35]

    Introducing gpt-5

    OpenAI. Introducing gpt-5. https://openai.com/index/introducing-gpt-5/, 2025 a

    2025

  27. [36]

    Introducing openai o3 and o4-mini

    OpenAI. Introducing openai o3 and o4-mini. https://openai.com/index/introducing-o3-and-o4-mini/, 2025 b

    2025

  28. [37]

    Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients

    Brenden K Petersen, Mikel Landajuela, T Nathan Mundhenk, Claudio P Santiago, Soo K Kim, and Joanne T Kim. Deep symbolic regression: Recovering mathematical expressions from data via risk-seeking policy gradients. arXiv preprint arXiv:1912.04871, 2019

    1912 arXiv

  29. [38]

    Large language models are zero shot hypothesis proposers

    Biqing Qi, Kaiyan Zhang, Haoxiang Li, Kai Tian, Sihang Zeng, Zhang-Ren Chen, and Bowen Zhou. Large language models are zero shot hypothesis proposers. arXiv preprint arXiv:2311.05965, 2023

    2023 arXiv

  30. [39]

    Towards scientific discovery with generative ai: Progress, opportunities, and challenges

    Chandan K Reddy and Parshin Shojaee. Towards scientific discovery with generative ai: Progress, opportunities, and challenges. 39 0 (27): 0 28601--28609, 2025

    2025

  31. [40]

    Mathematical discoveries from program search with large language models

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from program search with large language models. volume 625, pp.\...

    2024

  32. [41]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  33. [42]

    Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K. Reddy. Llm-sr: Scientific equation discovery via programming with large language models, March 2025 a

    2025

  34. [43]

    Doan, and Chandan K

    Parshin Shojaee, Ngoc-Hieu Nguyen, Kazem Meidani, Amir Barati Farimani, Khoa D. Doan, and Chandan K. Reddy. Llm-srbench: A new benchmark for scientific equation discovery with large language models, June 2025 b

    2025

  35. [44]

    Nonlinear dynamics and chaos: with applications to physics, biology, chemistry, and engineering (studies in nonlinearity)

    Steven H Strogatz. Nonlinear dynamics and chaos: with applications to physics, biology, chemistry, and engineering (studies in nonlinearity). 1, 2001

    2001

  36. [45]

    Thinking with images for multimodal reasoning: Foundations, methods, and future frontiers

    Zhaochen Su, Peng Xia, Hangyu Guo, Zhenhua Liu, Yan Ma, Xiaoye Qu, Jiaqi Liu, Yanshu Li, Kaide Zeng, Zhengyuan Yang, et al. Thinking with images for multimodal reasoning: Foundations, methods, and future frontiers. arXiv preprint arXiv:2506.23918, 2025

    2025 arXiv

  37. [46]

    Symbolic physics learner: Discovering governing equations via monte carlo tree search

    Fangzhao Sun, Yang Liu, Jianxun Hao, and George Em Karniadakis. Symbolic physics learner: Discovering governing equations via monte carlo tree search. In International Conference on Learning Representations, 2023

    2023

  38. [47]

    Ai feynman: A physics-inspired method for symbolic regression, April 2020

    Silviu-Marian Udrescu and Max Tegmark. Ai feynman: A physics-inspired method for symbolic regression, April 2020

    2020

  39. [48]

    Marco Virgolin and Solon P. Pissis. Symbolic regression is NP -hard. Transactions on Machine Learning Research, 2022

    2022

  40. [49]

    Scientific discovery in the age of artificial intelligence

    Hanchen Wang, Tianfan Fu, Yuanqi Du, Wenhao Gao, Kexin Huang, Ziming Liu, Payal Chandak, Shengchao Liu, Peter Van Katwyk, Andreea Deac, et al. Scientific discovery in the age of artificial intelligence. Nature, 620 0 (7972): 0 47--60, 2023 a

    2023

  41. [50]

    Physunibench: An undergraduate-level physics reasoning benchmark for multimodal models

    Lintao Wang, Encheng Su, Jiaqi Liu, Pengze Li, Peng Xia, Jiabei Xiao, Wenlong Zhang, Xinnan Dai, Xi Chen, Yuan Meng, et al. Physunibench: An undergraduate-level physics reasoning benchmark for multimodal models. arXiv preprint arXiv:2506.17667, 2025

    2025

  42. [51]

    Scimon: Scientific inspiration machines optimized for novelty

    Qingyun Wang, Doug Downey, Heng Ji, and Tom Hope. Scimon: Scientific inspiration machines optimized for novelty. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.\ 279--299, 2024

    2024

  43. [52]

    Hypothesis search: Inductive reasoning with language models

    Ruocheng Wang, Eric Zelikman, Gabriel Poesia, Yewen Pu, Nick Haber, and Noah D Goodman. Hypothesis search: Inductive reasoning with language models. arXiv preprint arXiv:2309.05660, 2023 b

    2023 arXiv

  44. [53]

    Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks

    Zhaofeng Wu, Linlu Qiu, Alexis Ross, Ekin Aky \"u rek, Boyuan Chen, Bailin Wang, Najoung Kim, Jacob Andreas, and Yoon Kim. Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks. Association for Computational Linguistics, 2024

    2024

  45. [54]

    Grok 3 beta — the age of reasoning agents

    xAI. Grok 3 beta — the age of reasoning agents. https://x.ai/news/grok-3, 2025

    2025

  46. [55]

    Vision-language models for vision tasks: A survey

    Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024 a

    2024

  47. [56]

    A comprehensive survey of scientific large language models and their applications in scientific discovery

    Yu Zhang, Xiusi Chen, Bowen Jin, Sheng Wang, Shuiwang Ji, Wei Wang, and Jiawei Han. A comprehensive survey of scientific large language models and their applications in scientific discovery. arXiv preprint arXiv:2406.10833, 2024 b

    2024 arXiv

  48. [57]

    Can gpt-4 perform neural architecture search? arXiv preprint arXiv:2304.10970, 2023 a

    Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie. Can gpt-4 perform neural architecture search? arXiv preprint arXiv:2304.10970, 2023 a

    2023 arXiv

  49. [58]

    Large language models for scientific synthesis, inference and explanation

    Yizhen Zheng, Huan Yee Koh, Jiaxin Ju, Anh TN Nguyen, Lauren T May, Geoffrey I Webb, and Shirui Pan. Large language models for scientific synthesis, inference and explanation. arXiv preprint arXiv:2310.07984, 2023 b

    2023 arXiv

  50. [59]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  51. [60]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  52. [61]

    and Reddy, Chandan K

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    2025 doi arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.