SciΠvot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents
Paper • 2610.21082 • Published • 11
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Please tell us who you are. Access is granted immediately.
Log in or Sign Up to review the conditions and access this dataset content.
SciPIvot
Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents
Paper · Code · Landing page
SciPIvot has 248 cases. In each one an evaluator chooses between the plan an LLM research agent proposed and the plan a senior researcher corrected it to. In the project condition the evaluator also reads the project record as it stood just before the correction. This dataset holds those project records.
Contents
| Path | Contents |
|---|---|
questions.jsonl |
The 248 cases: id, options (A, B), reference_answer, and snapshot, the record the case is asked against |
snapshots/<project>/<commit>.tar.zst |
230 project records from 26 projects, one per commit; some cases share one |
snapshots.jsonl |
One row per record: sha256 of the archive, tree_sha256 and files of the unpacked tree, sizes, and skipped (links that point outside the project, which the evaluated agents did not see either) |
A record is the project's git tree at the case's commit with no .git, so it holds no history and nothing after the correction. It is byte-for-byte what was mounted read-only at /project for the models in the paper. All 230 records come to 21 GB compressed.
Use
The GitHub repository has the scripts:
git clone https://github.com/AutoResearch-Factory/SciPIvot && cd SciPIvot
python3 -m venv .venv && . .venv/bin/activate # Python 3.12+
pip install huggingface_hub zstandard
hf auth login # after accepting the terms on this page
python scripts/fetch_records.py --id case_000 --out records
python scripts/fetch_records.py --all --out records --remove-after-verify # check all 230
python scripts/run_project.py --id case_000 --engine claude --model <model> --records records --out runs.jsonl
python scripts/score.py runs.jsonl
fetch_records.py checks every archive against sha256 and every unpacked tree against tree_sha256.
Citation
@misc{scipivot2026,
title = {{SciPIvot}: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents},
author = {Sun, Youran and Ren, Xingyu and Yi, Chugang and Zhang, Kejia and Wu, Di and Cao, Yingkang and Zhou, Allen and Ye, Jianping and Liu, Suying and Guo, Jiaxuan and Du, Jianda and Ye, Weicheng and Liu, Kecheng and Zhang, Hefan and Wang, Haixin and Xiao, Jie and Yang, Eric and Yang, Haizhao},
year = {2026},
eprint = {2610.21082},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.21082}
}
- Downloads last month
- 30