CausalPhys - Intervention: leaderboard
CausalPhys (2026) tests vision-language models on physical reasoning questions over real-world images and videos, each paired with an expert causal graph of objects, attributes and events. This board is answer accuracy in percent on the Intervention domain, reasoning under explicit changes of viewpoint, space, time or action; higher is better.
Metric: Answer accuracy (0-1 scaled to %) on the Intervention domain (reasoning under explicit manipulations such as viewpoint, spatial or temporal changes) of the image- and video-based questions with verified answers over real-world physical scenes; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 57.07 | #741 |
| 2 | InternVL3-78B | 57.07 | #738 |
| 3 | Claude Sonnet 4 | 56.72 | #400 |
| 4 | Gemini 2.5 Flash | 55.67 | #453 |
| 5 | GPT-4o Mini | 51.31 | #1623 |
| 6 | Mistral Small 3.2 | 48.52 | #1591 |
| 7 | Phi-4 Multimodal Instruct | 48.52 | #3882 |
| 8 | Qwen 2 VL 7B | 47.64 | #2261 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=causalphys-intervention · How It Works · Data refreshed daily, snapshot 2026-11-13.