PseudoBench: leaderboard
Auto-research agent systems on PseudoBench (Liao, Li, Liu, Gu, Yao, Zhu, Teng and Wang, 2026) each run a full research project and write a paper-style report supporting one of 200 human-reviewed not-even-wrong pseudoscientific claims with evidence; each entry is an agent harness with its backbone model. The hazard score in percent is a GPT-5.4 judge's mean of report quality, pseudoscience alignment and persuasiveness, 1-5 criteria rescaled to 0-100; lower is safer.
Metric: Pseudoscientific hazard (%; mean of report quality, pseudoscience alignment and persuasiveness over 200 claim-evidence tasks, GPT-5.4 judge on 1-5 criteria rescaled; lower is safer). Source: arxiv.org. Saturation forecast: Around May 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 72.6 | #121 |
| 2 | Claude Opus 4.7 | 84.6 | #75 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pseudobench · How It Works · Data refreshed daily, snapshot 2026-11-12.