MVX-Bench - Defeasible Entailment: leaderboard
Metric: Accuracy (%) on the 150 Defeasible Entailment questions (decide whether an update video strengthens or weakens a hypothesis), MVX-Bench's multi-video multiple-choice questions (each pairs a question with several videos; 1,442 questions over 4,255 videos repurposed from 11 annotated video datasets), end-to-end inference at temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 4B (IT) | 56 | #971 |
| 2 | Gemma 3 27B (IT) | 52.7 | #509 |
| 3 | GPT-5.2 | 52.7 | #105 |
| 4 | GPT-4o | 52 | #333 |
| 5 | Gemma 3 12B (IT) | 52 | #655 |
| 6 | Qwen 2.5 VL 32B Instruct | 51.3 | #443 |
| 7 | Phi-4 Multimodal Instruct | 48 | #896 |
| 8 | Qwen 3.6 27B (Thinking) | 47.3 | #163 (Qwen 3.6 27B) |
| 9 | Qwen 3.6 27B (Non-reasoning) | 47.3 | #163 (Qwen 3.6 27B) |
| 10 | Pixtral-12B | 46.7 | #795 |
| 11 | Qwen 2.5 VL 7B Instruct | 44.7 | #643 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=mvx-bench-defeasible-entailment · How It Works · Data refreshed daily, snapshot 2026-10-11.