CausalPhys - Description Correctness: 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 the share in percent of attribute and event descriptions that the model rationale states consistently with the annotation, checked by a GPT-4o judge; higher is better.
Metric: Description Correctness (0-1 scaled to %), the share of attribute and event descriptions the rationale states consistently with the annotation, averaged over the four domains; the model writes a rationale and an answer; a GPT-4o judge makes binary checks of the rationale against the expert-annotated causal graph (typed object, attribute and event nodes with directed dependencies); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-78B | 47.2 | #738 |
| 2 | GPT-4o | 43.97 | #741 |
| 3 | Gemini 2.5 Flash | 43.08 | #453 |
| 4 | Claude Sonnet 4 | 40.37 | #400 |
| 5 | GPT-4o Mini | 39.94 | #1623 |
| 6 | Mistral Small 3.2 | 36.69 | #1591 |
| 7 | Phi-4 Multimodal Instruct | 35.01 | #3882 |
| 8 | Qwen 2 VL 7B | 33.08 | #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-description-correctness · How It Works · Data refreshed daily, snapshot 2026-11-13.