| title | SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones? |
|---|---|
| authors | Sy-Tuyen Ho, Minghui Liu, Huy Nghiem, Furong Huang |
| date | 2026-05-28 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.30329 |
| affiliation | Furong Huang's Lab at UMD |
| upvotes | 4 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | Autonomous AI research agents aim to accelerate scientific discovery by automating the research pipeline, from hypothesis generation to peer review. However, existing benchmarks rarely test a fundamental bottleneck: whether Large Language Models can judge the methodological viability of a research idea before expending time and computational resources. We introduce SoundnessBench, a curated benchmark of 1,099 machine-learning research proposals reconstructed from ICLR submissions, labeled with reviewer soundness sub-scores, and audited against source papers. SoundnessBench should be interpreted as a benchmark for recoverable proposal-stage soundness rather than exact prediction of full-paper review outcomes. Across 12 frontier LLMs, we find a pervasive optimism bias: under standard prompting, models frequently rate low-soundness proposals as sound, while aggressive prompting largely shifts errors from false positives to false negatives. Additional controls for public-corpus contamination, paper-identifying phrases, surface features, and human audit quality suggest that this behavior is not explained by a single confounder. Our results indicate that current LLMs are not yet reliable as standalone first-gate evaluators for scientific rigor. |
cite
@misc{ho2026soundnessbench,
title={SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?},
author={Sy-Tuyen Ho and Minghui Liu and Huy Nghiem and Furong Huang},
year={2026},
eprint={2605.30329},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.30329}
}