Elamur Research⌕
☰
home / papers / 2026 / SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?

SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?

Paper · 2026 · Sy-Tuyen Ho, Minghui Liu, Huy Nghiem et al.

Summary

AI "scientist" agents can now draft hypotheses, code, and papers. Current benchmarks, however, mostly test execution and skip an earlier question: can a model spot a methodologically weak research plan before compute is spent on it? SoundnessBench targets this pre-execution triage step for machine learning. The authors started from ICLR submissions and their reviews. They kept papers where reviewers agreed strongly and labeled each one high- or low-soundness using the reviewers' soundness sub-scores. Each paper was turned into a near-verbatim proposal with the results hidden, and the extraction was checked against the source PDF through retrieval-based verification of atomic claims.

Twelve frontier LLMs were tested on these proposals. Under ordinary prompts they were strongly optimistic and approved most weak proposals. A harsher prompt did not fix this; it swung the models toward rejecting good proposals instead. Robustness checks did not make the pattern disappear. These included contamination, identifier, surface-feature and injected-flaw analyses, plus a human audit. The authors conclude that prompting alone does not make LLMs dependable first-gate reviewers of research rigor.

Key points

  • The dataset contains 1,099 proposals (458 low-soundness, 641 high-soundness). They were selected from 35,209 ICLR submissions and 137,940 reviews spanning 2022–2026 and cover 16 ML subfields.
  • Labeling rules:
    • Mean reviewer soundness ≥ 3 counts as high; ≤ 2 counts as low, and middle scores are dropped.
    • Only papers with mean reviewer confidence ≥ 3/5 and soundness std < 15% are kept.
  • With standard prompting, the 12 models had a mean false-positive rate of 74.0%. An aggressive prompt cut this to 19.9%, but recall on high-soundness proposals fell to 36.1%.
  • Robustness checks included:
    • an ICLR 2026-only split to limit contamination
    • removal of paper-identifying phrases
    • surface-feature, year, subfield and writing-quality slices
    • adversarial flaw injection
    • a preliminary human audit
  • The dataset is published on Hugging Face as hosytuyen/SoundnessBench. The authors frame it as a test of how much soundness can be recovered from a proposal alone, not as a predictor of full-paper review outcomes.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of Maryland, College ParkCollege Park · United States149 works · 29,191 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 19,503 worksArtificial Intelligence · 18,805 worksArts and Humanities · 17,959 worksLanguage and Linguistics · 17,789 worksSocial Sciences · 2,009 worksSafety Research · 665 worksHistory and Philosophy of Science · 83 works

Cite

Cite as Sy-Tuyen Ho, Minghui Liu, Huy Nghiem and Furong Huang. 2026. SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?. arXiv:2605.30329.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2605.30329 · arXiv 2605.30329

@misc{ho2026soundnessbench,
  title = {{SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?}},
  author = {Sy-Tuyen Ho, Minghui Liu, Huy Nghiem and Furong Huang},
  year = {2026},
  doi = {10.48550/arxiv.2605.30329},
  eprint = {2605.30329},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2605.30329}
}

References

21 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, et al. Mle-bench: Evaluating machine learning agents on machine learning engineering. arXiv preprint arXiv:2410.07095, 2024.
  2. Katherine Elkins and Jon Chun. Syntactic framing fragility: An audit of robustness in llm ethical decisions. arXiv preprint arXiv:2601.09724, 2025. — [arXiv:2601.09724](https://arxiv.org/abs/2601.09724) find
  3. Longling Geng, Andy Ouyang, Theodore Wu, Daphne Barretto, Matthew John Hayes, Rachael Cooper, Yuqiao Zeng, Sameer Vijay, Gia Ancone, Ankit Rai, et al. Causalt5k: Diagnosing and informing refusal for trustworthy causal reasoning of skepticism, sycophancy, detection-correction, and rung collapse. arXi… — [arXiv:2602.08939](https://arxiv.org/abs/2602.08939) find
  4. Yehua Huang, Penglei Sun, Zebin Chen, Zhenheng Tang, and Xiaowen Chu. Omnireview: A large-scale benchmark and llm-enhanced framework for realistic reviewer recommendation. arXiv preprint arXiv:2602.08896, 2026. — [arXiv:2602.08896](https://arxiv.org/abs/2602.08896) find
  5. Bo Jiang. Hindsight: Evaluating research idea generation via future impact. arXiv preprint arXiv:2603.15164, 2026. — [arXiv:2603.15164](https://arxiv.org/abs/2603.15164) find
  6. Andrej Karpathy. autoresearch. https://github.com/karpathy/autoresearch , 2026. GitHub repository, accessed 2026-03-27. ↗
  7. Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist: Towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292, 2024.
  8. Alisia Lupidi, Bhavul Gauri, Thomas Simon Foster, Bassel Al Omari, Despoina Magka, Alberto Pepe, Alexis Audran-Reiss, Muna Aghamelu, Nicolas Baldwin, Lucia Cipolina-Kun, et al. Airs-bench: a suite of tasks for frontier ai research science agents. arXiv preprint arXiv:2602.06855, 2026. — [arXiv:2602.06855](https://arxiv.org/abs/2602.06855) find
  9. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 202…
  10. Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Michael Moor, Zicheng Liu, and Emad Barsoum. Agent laboratory: Using llm agents as research assistants. Findings of the Association for Computational Linguistics: EMNLP 2025, pages 5977–6043, 2025.
  11. Tim Schopf and Michael Färber. Is this idea novel? an automated benchmark for judgment of research ideas. arXiv preprint arXiv:2603.10303, 2026.
  12. Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R Johnston, et al. Towards understanding sycophancy in language models. arXiv preprint arXiv:2310.13548, 2023.
  13. Chenglei Si, Tatsunori Hashimoto, and Diyi Yang. The ideation-execution gap: Execution outcomes of llm-generated versus human research ideas. arXiv preprint arXiv:2506.20803, 2025a.
  14. Chenglei Si, Diyi Yang, and Tatsunori Hashimoto. Can LLMs generate novel research ideas? a large-scale human study with 100+ NLP researchers. In The Thirteenth International Conference on Learning Representations, 2025b. URL https://openreview.net/forum?id=M23dTGWCZy . ↗
  15. Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, Johannes Heidecke, Amelia Glaese, and Tejal Patwardhan. Paperbench: Evaluating AI’s ability to replicate AI research. In Forty-second International Confere… — PaperBench: Evaluating AI's Ability to Replicate AI Research
  16. Nitya Thakkar, Mert Yuksekgonul, Jake Silberg, Animesh Garg, Nanyun Peng, Fei Sha, Rose Yu, Carl Vondrick, and James Zou. Can llm feedback enhance review quality? a randomized study of 20k reviews at iclr 2025. arXiv preprint arXiv:2504.09737, 2025.
  17. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. Fever: a large-scale dataset for fact extraction and verification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1…
  18. Jingqi Tong, Mingzhe Li, Hangcheng Li, Yongzhuo Yang, Yurong Mou, Weijie Ma, Zhiheng Xi, Hongji Chen, Xiaoran Liu, Qinyuan Cheng, et al. Ai can learn scientific taste. arXiv preprint arXiv:2603.14473, 2026. — [arXiv:2603.14473](https://arxiv.org/abs/2603.14473) find
  19. Yunze Wu, Dayuan Fu, Weiye Si, Zhen Huang, Mohan Jiang, Keyu Li, Shijie Xia, Jie Sun, Tianze Xu, Xiangkun Hu, et al. Innovatorbench: Evaluating agents’ ability to conduct innovative llm research. arXiv preprint arXiv:2510.27598, 2025.
  20. Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist-v2: Workshop-level automated scientific discovery via agentic tree search. arXiv preprint arXiv:2504.08066, 2025.
  21. Sungduk Yu, Man Luo, Avinash Madusu, Vasudev Lal, and Phillip Howard. Is your paper being reviewed by an llm? benchmarking ai text detection in peer review. arXiv preprint arXiv:2502.19614, 2025. — [arXiv:2502.19614](https://arxiv.org/abs/2502.19614) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ho-2026-soundnessbench-ai-scientist-really/