Under Review

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

Sy-Tuyen Ho, Minghui Liu, Huy Nghiem, Furong Huang

University of Maryland, College Park

Paper 🤗   Dataset Code BibTeX
1,099
Real/Validated proposals from ICLR 2022–2026
12
Frontier LLMs evaluated across open & closed-source
74%
Mean false-positive rate on low-soundness proposals
16
ML subfields covered — RL, CV, NLP & more

Leaderboard

Rank Model Eval Mode
1 Claude-Sonnet-4-6 Standard Prompt 62.8% 5.2% 94.8% 37.2% 94.8% 67.7% 79.0%
2 claude-opus-4-6 Standard Prompt 71.8% 2.8% 97.2% 28.2% 97.2% 64.8% 77.7%
3 Gemini-3.1-Pro Standard Prompt 74.0% 3.9% 96.1% 26.0% 96.1% 64.5% 77.2%
4 Gemini-3-Flash Standard Prompt 80.3% 1.9% 98.1% 19.7% 98.1% 63.1% 76.8%
5 Gemini-2.5-Pro Standard Prompt 86.2% 1.6% 98.4% 13.8% 98.4% 61.5% 75.7%
6 Qwen3.5-27B Standard Prompt 76.4% 7.5% 92.5% 23.6% 92.5% 62.9% 74.9%
7 GPT-5.4 Standard Prompt 35.4% 25.4% 74.6% 64.6% 74.6% 74.7% 74.6%
8 Qwen3.5-122B-A10B Standard Prompt 73.4% 9.4% 90.6% 26.6% 90.6% 63.4% 74.6%
9 GPT-4o Standard Prompt 94.5% 1.1% 98.9% 5.5% 98.9% 59.4% 74.2%
10 LLaMA-3.3-70B-Instruct Standard Prompt 98.0% 0.6% 99.4% 2.0% 99.4% 58.7% 73.8%
11 GPT-5.4-Mini Standard Prompt 49.3% 23.4% 76.6% 50.7% 76.6% 68.5% 72.3%
12 claude-opus-4-6 Aggressive Prompt 27.8% 33.6% 66.4% 72.2% 66.4% 73.9% 70.0%
13 LLaMA-3.3-70B-Instruct Aggressive Prompt 60.7% 23.9% 76.1% 39.3% 76.1% 63.7% 69.4%
14 Kimi-Linear-48B-A3B Standard Prompt 86.2% 15.4% 84.6% 13.8% 84.6% 57.8% 68.7%
15 Gemini-3-Flash Aggressive Prompt 23.6% 39.6% 60.4% 76.4% 60.4% 78.2% 68.1%
16 Gemini-2.5-Pro Aggressive Prompt 26.6% 44.9% 55.1% 73.4% 55.1% 74.3% 63.3%
17 Kimi-Linear-48B-A3B Aggressive Prompt 49.7% 53.0% 47.0% 50.3% 47.0% 57.0% 51.5%
18 Gemini-3.1-Pro Aggressive Prompt 10.9% 64.1% 35.9% 89.1% 35.9% 82.1% 49.9%
19 Qwen3.5-27B Aggressive Prompt 16.2% 67.6% 32.4% 83.8% 32.4% 73.8% 45.1%
20 GPT-4o Aggressive Prompt 12.9% 75.0% 25.0% 87.1% 25.0% 73.1% 37.2%
21 claude-sonnet-4-6 Aggressive Prompt 5.6% 81.2% 18.8% 94.4% 18.8% 81.5% 30.6%
22 Qwen3.5-122B-A10B Aggressive Prompt 4.4% 83.2% 16.8% 95.6% 16.8% 84.4% 28.1%
23 GPT-5.4-Mini Aggressive Prompt 0.0% 99.8% 0.2% 100.0% 0.2% 100.0% 0.3%
24 GPT-5.4 Aggressive Prompt 0.0% 100.0% 0.0% 100.0% 0.0% -- --

Abstract

TL;DR. Frontier LLMs are systematically over-optimistic when judging research soundness. They approve weak proposals 74% of the time. Aggressive prompting does not improve discrimination; it largely shifts errors from false positives to false negatives. Our results indicate that current LLMs are not yet reliable as standalone first-gate evaluators for scientific rigor.

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.

Contributions

01

SoundnessBench: A Large-Scale Pre-Execution Benchmark

1,099 research proposals spanning 8 years of top-tier ML submissions (ICLR 2022–2026) and 16 sub-disciplines grounded in real expert peer-review outcomes.

02

High-Precision Multi-Stage Curation Pipeline

Expert-agreement filtering (4,391 human reviews), atomic-claim auditing to prevent benchmark leakage, near-verbatim proposal extraction, and outcome masking ensuring each proposal is faithfully traceable to source evidence.

03

Quantifying the Optimism-Fragility Tradeoff

Empirical study of 12 frontier LLMs revealing systemic failure to identify methodological flaws. We characterize an "optimism bias" (74% FPR) that persists across model scales, families, and instruction-tuning stages.

04

Prompt Fragility as a Core Failure Mode

Aggressive prompting doesn't resolve optimism bias and does not improve discrimination; it largely shifts errors from false positives to false negatives.

Benchmark Construction

SoundnessBench is reconstructed from the ICLR public history. We processed over 35,209 initial submissions and 137,940 expert reviews to distill a high-signal subset with auditable, traceable labels.

SoundnessBench Curation Pipeline Flowchart
SoundnessBench pipeline: (1) collect ICLR papers with reviewer metadata and filter for high reviewer agreement; (2) derive high/low-soundness labels; (3) extract near-verbatim research proposal without revealing experimental results; (4) audit extraction fidelity with retrieve-then-verify atomic claims; and (5) assemble the final benchmark.

Dataset Statistics

The final benchmark contains 1,099 research proposals: 458 low-soundness and 641 high-soundness instances, with clear score separation between classes.

Dataset Distribution Graphs and Score Density Plots
SoundnessBench dataset statistics. The benchmark includes 1,099 proposals, including 458 low-soundness and 641 high-soundness instances. (a) Subfield distribution across papers reflects ICLR corpus composition. (b) Soundness score density shows separation between low-soundness (S < 2, mean = 1.77) and high-soundness (S > 3, mean = 3.22) groups, supporting the chosen label boundary. (c) Temporal coverage spans ICLR 2022–2026. (d) Low- and high-soundness pair-count statistics in SoundnessBench.

Main Results

A Consistent Optimism Bias Across All Model Families

Standard prompt confusion matrices for 12 frontier models
Confusion matrices under the standard prompt across 12 evaluated models. Main message: many models are over-optimistic by default. The mean false-positive rate on low-soundness proposals is 74.0% (9/12 models exceed 70%). This pattern appears across model families in this evaluation setting.

Aggressive Prompting Analysis

Aggressive Prompting Alone Cannot Fix the Bias

Aggressive prompt confusion matrices for 12 models
Confusion matrices under the aggressive prompt across 12 evaluated models. Main message: optimism bias often shifts toward over-conservatism. The mean false-positive rate on low-soundness proposals drops to 19.9% (10/12 models are below 30%), but recall on high-soundness proposals also drops to 36.1% (7/12 models are below 40%). This illustrates strong prompt sensitivity in scientific-soundness judgment for the evaluated models.
Key Takeaway 1

Frontier LLMs show a broad optimism bias: models frequently rate low-soundness proposals as sound under standard prompting.

Key Takeaway 2

The tested stricter prompt does not jointly improve discrimination on both classes; it mainly shifts errors from false positives to false negatives.

Key Takeaway 3

Taken together, results point to a prompt-sensitive capability limitation in proposal-stage soundness judgment: current LLMs are not yet reliable as standalone gatekeepers for scientific rigor.

Related Benchmarks

Prior benchmarks focus on execution outcomes or post-hoc review of completed papers. SoundnessBench is the only benchmark combining pre-execution evaluation, direct methodological-soundness judgment, and proposal-only input with verified ground truth.

Benchmark Stage Task Input Soundness GT Pre-Exec.
MLE-Bench Execution Engineering Task + Results ✗ ✗
PaperBench Execution Replication Full paper + Results ✗ ✗
InnovatorBench Execution Research loop Task + Results ✗ ✗
Si et al. (2025) Pre-exec. Novelty judge Proposal ✗ ✓
Hindsight Post-hoc Impact pred. Idea ✗ ✗
RINoBench Pre-exec. Novelty judge Idea + related works ✗ ✓
SoundnessBench (Ours) Pre-exec. Methodological soundness Proposal ✓ ✓

Citation

@misc{ho2026soundnessbenchaiscientistreally,
      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},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.30329}, 
}