NoisyCausal - Question Perturbation: leaderboard
Causal reasoning under structured noise (NoisyCausal; Xu and Fu, 2026), question perturbation condition: random 3-7 variable causal graphs are grounded in medicine, education or economics scenarios and the model answers an interventional, counterfactual or attributional question whose answer follows from the clean model, with a question whose assumptions contradict the structural model in the story. Accuracy in percent, higher is better.
Metric: Accuracy (%; synthetic structural-causal-model questions with a question whose assumptions contradict the structural model). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 48.5 | #996 |
| 2 | GPT-3.5 | 43.8 | #2430 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=noisycausal-question-perturbation · How It Works · Data refreshed daily, snapshot 2026-11-12.