Preprint · 2026

PRBench: End-to-end Paper Reproduction in Physics Research

Shi Qiu, Junyi Deng, Yiwei Deng, Haoran Dong, Jieyu Fu, Mao Li, Zeyu Li, Zhaolong Zhang, Huiwen Zheng, Leidong Bao, Anqi Lv, Zihan Mo, Yadi Niu, Yiyang Peng, Yu Tian, Yili Wang, Ziyu Wang, Zi-Yu Wang, Jiashen Wei, Liuheng Wu, Aoran Xue, Leyi Yang, Guanglu Yuan, Xiarui Zhan, Jingjun Zhang, Zifan Zheng, Pengfei Liu, Linrui Zhen, Kaiyang Li, Qichang Li, Ziheng Zhou, Guo-En Nian, Yunwei Xiao, Qing-Hong Cao, Linjie Dai, Xu Feng, Peng Gao, Ying Gu, Chang Liu, Jia Liu, Ming-xing Luo, Yan-Qing Ma, Liang-You Peng, Huichao Song, Shufeng Wang, Chenxu Wang, Tao Wang, Yi-Nan Wang, Chengyin Wu, Pengwei Zhao, Hua Xing Zhu

Abstract

AI agents powered by large language models exhibit strong reasoning and problem-solving capabilities, enabling them to assist scientific research tasks such as formula derivation and code generation. However, whether these agents can reliably perform end-to-end reproduction from real scientific papers remains an open question. We introduce PRBench, a benchmark of 30 expert-curated tasks spanning 11 subfields of physics. Each task requires an agent to comprehend the methodology of a published paper, implement the corresponding algorithms from scratch, and produce quantitative results matching the original publication. Agents are provided only with the task instruction and paper content, and operate in a sandboxed execution environment. All tasks are contributed by domain experts from over 20 research groups at the School of Physics, Peking University, each grounded in a real published paper and validated through end-to-end reproduction with verified ground-truth results and detailed scoring rubrics. Using an agentified assessment pipeline, we evaluate a set of coding agents on PRBench and analyze their capabilities across key dimensions of scientific reasoning and execution. The best-performing agent, OpenAI Codex powered by GPT-5.3-Codex, achieves a mean overall score of 34%. All agents exhibit a zero end-to-end callback success rate, with particularly poor performance in data accuracy and code correctness. We further identify systematic failure modes, including errors in formula implementation, inability to debug numerical simulations, and fabrication of output data. Overall, PRBench provides a rigorous benchmark for evaluating progress toward autonomous scientific research.

BibTeX Copy

@misc{qiu_prbench_2026,
	title = {{PRBench}: {End}-to-end {Paper} {Reproduction} in {Physics} {Research}},
	shorttitle = {{PRBench}},
	url = {http://arxiv.org/abs/2603.27646},
	doi = {10.48550/arXiv.2603.27646},
	abstract = {AI agents powered by large language models exhibit strong reasoning and problem-solving capabilities, enabling them to assist scientific research tasks such as formula derivation and code generation. However, whether these agents can reliably perform end-to-end reproduction from real scientific papers remains an open question. We introduce PRBench, a benchmark of 30 expert-curated tasks spanning 11 subfields of physics. Each task requires an agent to comprehend the methodology of a published paper, implement the corresponding algorithms from scratch, and produce quantitative results matching the original publication. Agents are provided only with the task instruction and paper content, and operate in a sandboxed execution environment. All tasks are contributed by domain experts from over 20 research groups at the School of Physics, Peking University, each grounded in a real published paper and validated through end-to-end reproduction with verified ground-truth results and detailed scoring rubrics. Using an agentified assessment pipeline, we evaluate a set of coding agents on PRBench and analyze their capabilities across key dimensions of scientific reasoning and execution. The best-performing agent, OpenAI Codex powered by GPT-5.3-Codex, achieves a mean overall score of 34\%. All agents exhibit a zero end-to-end callback success rate, with particularly poor performance in data accuracy and code correctness. We further identify systematic failure modes, including errors in formula implementation, inability to debug numerical simulations, and fabrication of output data. Overall, PRBench provides a rigorous benchmark for evaluating progress toward autonomous scientific research.},
	urldate = {2026-11-06},
	publisher = {arXiv},
	author = {Qiu, Shi and Deng, Junyi and Deng, Yiwei and Dong, Haoran and Fu, Jieyu and Li, Mao and Li, Zeyu and Zhang, Zhaolong and Zheng, Huiwen and Bao, Leidong and Lv, Anqi and Mo, Zihan and Niu, Yadi and Peng, Yiyang and Tian, Yu and Wang, Yili and Wang, Ziyu and Wang, Zi-Yu and Wei, Jiashen and Wu, Liuheng and Xue, Aoran and Yang, Leyi and Yuan, Guanglu and Zhan, Xiarui and Zhang, Jingjun and Zheng, Zifan and Liu, Pengfei and Zhen, Linrui and Li, Kaiyang and Li, Qichang and Zhou, Ziheng and Nian, Guo-En and Xiao, Yunwei and Cao, Qing-Hong and Dai, Linjie and Feng, Xu and Gao, Peng and Gu, Ying and Liu, Chang and Liu, Jia and Luo, Ming-xing and Ma, Yan-Qing and Peng, Liang-You and Song, Huichao and Wang, Shufeng and Wang, Chenxu and Wang, Tao and Wang, Yi-Nan and Wu, Chengyin and Zhao, Pengwei and Zhu, Hua Xing},
	month = mar,
	year = {2026},
	note = {arXiv:2603.27646 [cs.CL]},
	keywords = {Computer Science - Computation and Language, High Energy Physics - Lattice, High Energy Physics - Phenomenology, Physics - Computational Physics, Physics - Optics},
}