PRBench: End-to-end Paper Reproduction in Physics Research
Shi Qiu, Jun Deng, Yiwei Deng, Haoran Dong, Jie Fu, Mao Li, Zeyu Li, Zhao Zhang, Huiwen Zheng, Lei Bao, Anqiang Lv, Zihan Mo, Yadi Niu, Yiyang Peng, Yu Tian, Yilin Wang, Ziyu Wang, Ziyu Wang, Jiashen Wei, Liu Wu, Ao Xue, Leyi Yang, Guanglu Yuan, Xia Zhan, Jingjun Zhang, Zifan Zheng, Pengfei Liu, Li Zhen, Kaiyang Li, Qichang Li, Ziheng Zhou, Guokui Nian, Yu-Xiao Xiao, Qing-Hong Cao, L. Dai, Xu Feng, Pengyu Gao, Ying Gu, Chang Liu, Jia Liu, M. Luo, Yan-Qing Ma, Liangchen Peng, Huichao Song, Shufeng Wang, Chenxu Wang, Tao Wang, Yi-Nan Wang, C. Wu, Peng Zhao, H. Zhu · 2026 · arXiv
TL;DR
This work evaluates a set of coding agents on PRBench and analyzes their capabilities across key dimensions of scientific reasoning and execution, identifying systematic failure modes, including errors in formula implementation, inability to debug numerical simulations, and fabrication of output data.
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.
Editorial notes
Introduces PRBench, a benchmark of 30 expert-curated end-to-end physics paper-reproduction tasks across 11 subfields, where the agent must comprehend a paper, implement its algorithms from scratch, and produce quantitative results matching the original publication. The best-performing agent (OpenAI Codex powered by GPT-5.3-Codex) achieves only 34% mean overall score, with a zero end-to-end callback success rate. Authors identify systematic failure modes including errors in formula implementation, inability to debug numerical simulations, and fabrication of output data — all directly on-model for SAICA-KG. Relates to SAICA-KG failure modes: fabrication, logic_error, incomplete_execution. Cite when arguing that long-horizon reproducibility is currently the hardest stress test for coding agents.