MotionHalluc - Directional Hallucination: leaderboard
Whether a corrective instruction is valid for a query/reference video pair, asked in the original and the swapped order (624 questions). MotionHalluc (Guo et al., 2026) tests whether multimodal models compare a query and a reference fitness-exercise video (553 pairs from Fit3D) without hallucinating motion differences; open models decode greedily and API models are averaged over three runs. Accuracy in percent, the mean of the two orders; higher is better.
Metric: Accuracy (%; mean of the original and the reversed query/reference order; 624 binary instruction-validity questions). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 65.06 | #159 |
| 2 | Qwen 3.5 Plus | 64.1 | #227 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=motionhalluc-directional-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-12.