MoHallBench - Sequential Inference Hallucination Rate (Generative): leaderboard
Share of free-form video descriptions that claim a withheld terminal action, such as a diver entering the water, on clips cut before it happens; judged by an LLM entailment check with manual review, from MoHallBench (2026). Percent, lower is better.
Metric: Hallucination rate (%; share of free-form descriptions of truncated clips that state or entail the withheld terminal action, judged by a DeepSeek entailment judge with manual review; lower is better). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Molmo2-8B | 43.3 | #1483 |
| 2 | Qwen 2.5 VL 7B Instruct | 56.8 | #1833 |
| 3 | Qwen 3 VL 8B Instruct | 57 | #981 |
| 4 | InternVL3-8B | 66.2 | #1688 |
| 5 | Qwen 2 VL 7B Instruct | 66.8 | #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-sequential-inference-hallucination-rate-generative · How It Works · Data refreshed daily, snapshot 2026-11-13.