SoundnessBench: leaderboard
SoundnessBench (Ho, Liu, Nghiem and Huang, University of Maryland, 2026) gives a model a results-masked research proposal reconstructed from an ICLR 2022-2026 submission and asks it to justify and then classify the proposal's methodological soundness as high or low, against labels from high-agreement reviewer soundness sub-scores (1,099 proposals, 641 high and 458 low). The score is macro-F1 in percent over the two classes under the paper's standard prompt; higher is better.
Metric: Macro-F1 (%; high vs low soundness of 1,099 results-masked ICLR proposals, standard prompt). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 69.7 | #121 |
| 2 | Claude Sonnet 4.6 | 65.3 | #158 |
| 3 | GPT-5.4 Mini | 63.8 | #389 |
| 4 | Claude Opus 4.6 | 60.5 | #100 |
| 5 | Gemini 3.1 Pro (Preview) | 58.4 | #91 |
| 6 | Qwen 3.5 122B A10B | 56.3 | #329 |
| 7 | Qwen 3.5 27B | 55.1 | #358 |
| 8 | Gemini 3 Flash | 54.5 | #159 |
| 9 | Gemini 2.5 Pro | 49.8 | #262 |
| 10 | GPT-4o | 42.3 | #741 |
| 11 | Llama 3.3 70B Instruct | 38.9 | #1353 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=soundnessbench · How It Works · Data refreshed daily, snapshot 2026-11-12.