MoHallBench - Semantic Similarity Hallucination: leaderboard
Tests whether video LLMs confuse actions that share a purpose but differ physically, such as CPR and the Heimlich manoeuvre, using bias-aware paired yes/no queries; from MoHallBench (2026). Q-PairAcc in percent, higher is better.
Metric: Q-PairAcc (%; share of adversarial pairs in which the model answers both the positive and the negative (reversed) yes/no query correctly; 32 uniformly sampled frames, temperature 0; semantically related but physically distinct actions). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B Instruct | 39.3 | #981 |
| 2 | Molmo2-8B | 27 | #1483 |
| 3 | InternVL3-8B | 25.8 | #1688 |
| 4 | Qwen 2.5 VL 7B Instruct | 20.7 | #1833 |
| 5 | Qwen 2 VL 7B Instruct | 16.4 | #2553 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mohallbench-semantic-similarity-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-13.