CausalPhys: 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. The board is answer accuracy in percent over the Perception, Anticipation, Intervention and Goal-Orientation domains, weighted by items; higher is better.
Metric: Answer accuracy (0-1 scaled to %) over all four domains (Perception, Anticipation, Intervention, Goal-Orientation; 3,062 image- and video-based questions with verified answers over real-world physical scenes), item-weighted; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-78B | 59.24 | #738 |
| 2 | GPT-4o | 58.88 | #741 |
| 3 | Gemini 2.5 Flash | 57.25 | #453 |
| 4 | GPT-4o Mini | 56.3 | #1623 |
| 5 | Claude Sonnet 4 | 54.28 | #400 |
| 6 | Phi-4 Multimodal Instruct | 51.99 | #3882 |
| 7 | Qwen 2 VL 7B | 50.65 | #2261 |
| 8 | Mistral Small 3.2 | 46.11 | #1591 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=causalphys · How It Works · Data refreshed daily, snapshot 2026-11-13.