PhySciBench

PhySciBench is a benchmark for evaluating deep-research capabilities in the physical sciences, introduced in "Deep Research in Physical Sciences: A Multi-Agent Framework and Comprehensive Benchmark" (arXiv:2606.18648).

• 📖 Paper: https://arxiv.org/abs/2606.18648
• 💻 Code & evaluation: https://github.com/yigengjiang/physci-deepresearch

Overview

PhySciBench comprises 200 expert-curated questions (a single test split), balanced between physics and chemistry, spanning six task categories (the type field) that reflect real-world scientific workflows:

• multimodal-qa — perception and reasoning over scientific figures
• long-context-qa — synthesis across full documents and supplementary materials
• structured-information-extraction — schema-conformant parsing into JSON/CSV
• scientific-reasoning — multi-step, principle-grounded derivation
• experimental-design — procedurally complete synthesis/characterization protocols
• code-generation — executable computational implementations

State-of-the-art systems struggle: the strongest baseline, Gemini Deep Research, reaches only 33.5% accuracy.

Files

• physcibench.json — 200 records (the test split; the scorable metadata).
• files/ — referenced figures and source PDFs (141 files). The official scorer is fully functional on physcibench.json alone.

Record schema

Field	Description
id	Unique id, e.g. physci-001
question	The question text
answer	Ground-truth answer
category	Reporting label (long-form-answer / atomic-answer)
type	Task category (one of the six above)
files	Referenced figure/PDF filenames under files/
rubrics	Scoring rubric (for rubric-graded items; null otherwise)

Usage

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download("littletreee/PhySciBench", "physcibench.json", repo_type="dataset")
data = json.load(open(path))
print(len(data), "records")

For the official LLM-as-judge evaluation pipeline (predictions.jsonl → metrics.json), see the GitHub repository.

License & usage

The evaluation code (on GitHub) is licensed under Apache-2.0.

PhySciBench is only used for academic research. Commercial use in any form is prohibited. The copyright of all third-party materials in files/ (papers, figures, tables, excerpts, datasets, and supplementary materials) belongs to their original owners and remains under their original copyrights and licenses; these are not covered by the PhySciBench license unless explicitly stated. If there is any infringement in PhySciBench, please email yigengjiang@gmail.com and we will remove it immediately. Without prior approval, you cannot distribute, publish, copy, disseminate, or modify PhySciBench in whole or in part. You must strictly comply with the above restrictions.

Citation

If you find our work helpful for your research, please consider citing our work.

@article{jiang2026physcidr,
  title   = {Deep Research in Physical Sciences: A Multi-Agent Framework and Comprehensive Benchmark},
  author  = {Jiang, Yigeng and Yang, Tengchao and Cui, Taoyong and Wan, Jiaxing and Wang, Yuan and Wang, Weida and Liu, Zhiyu and Peng, Chuyi and Luo, Binzhao and Gao, Maoli and Huang, Huaihai and Zeng, Yuqianer and Zheng, Ziyang and Huang, Dongchen and Chen, Chao and Liu, Zichao and Shen, Weiping and Pu, Shuchen and Zhou, Siyu and Ma, Runmin and Hu, Yusong and Chao, Fei and Zhang, Bo and Zheng, Xiawu and Wang, Zifu and Bai, Lei and Cai, Yunqi and Zhang, Shufei},
  journal = {arXiv preprint arXiv:2606.18648},
  year    = {2026}
}