SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?
Summary
SoundnessBench is a benchmark of 1,099 machine-learning research proposals that evaluates LLMs' ability to assess methodological validity, finding a pervasive optimism bias in current models.
View Cached Full Text
Cached at: 06/01/26, 11:20 AM
Paper page - SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?
Source: https://huggingface.co/papers/2605.30329
Abstract
SoundnessBench evaluates large language models’ ability to assess the methodological validity of machine learning research proposals, revealing persistent optimism bias in current models.
Autonomous AI research agentsaim to accelerate scientific discovery by automating the research pipeline, fromhypothesis generationtopeer review. However, existing benchmarks rarely test a fundamental bottleneck: whetherLarge Language Modelscan judge the methodological viability of a research idea before expending time and computational resources. We introduce SoundnessBench, a curated benchmark of 1,099machine-learning research proposalsreconstructed fromICLR submissions, labeled withreviewer soundnesssub-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.
View arXiv pageView PDFProject pageGitHub0Add to collection
Get this paper in your agent:
hf papers read 2605\.30329
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.30329 in a model README.md to link it from this page.
Datasets citing this paper1
#### hosytuyen/SoundnessBench Viewer• Updated2 days ago • 1.1k • 221
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.30329 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI
This paper introduces MLS-Bench, a benchmark designed to assess whether AI systems can invent generalizable and scalable machine learning methods rather than just performing engineering tuning.
Are LLMs Ready for Scientific Discovery? A Capability-Oriented Benchmark for AI Scientists
Introduces SDABench, a benchmark evaluating LLMs on six scientific analysis capabilities across five domains, finding models struggle with tasks requiring assumption selection and mechanistic reasoning.
Benchmarking the Benchmarks: Testing the Predictive Validity of Commonsense Benchmarks
This paper tests whether commonsense benchmark scores predict real-world downstream task performance by evaluating 23 LLMs across four benchmarks and their reworked variants, finding that revisions preserve rankings but only offer task-dependent predictive validity.
On the Limits of LLM-as-Judge for Scientific Novelty Assessment
This paper introduces RQ-Bench, a benchmark to evaluate LLMs' ability to assess the novelty of scientific research questions. It finds that LLM judges consistently rate generated questions as more novel than human experts do, raising concerns about the reliability of using LLMs for scientific novelty evaluation.
Good Benchmarks
This paper from July 2026 defines principles for designing effective benchmark tasks for AI agents, drawing on experience from the Terminal Bench project. Good tasks are correct, solvable, verifiable, well-specified, and hard for interesting reasons, with an emphasis on real-world relevance and outcome-based verification.