ClinHallu - Reasoning Integration Hallucination: leaderboard
The reasoning integration hallucination rate on ClinHallu (2026): how often a model reasons to unsupported conclusions even when given the correct visual findings and medical knowledge, in percent; lower is better. ClinHallu (2026) pairs medical VQA items from four datasets with validated reasoning traces and averages over the four datasets.
Metric: Reasoning integration hallucination rate (%): share of items whose reasoning stage, generated after the reference visual and knowledge stages are supplied, the judge marks as hallucinated, macro-averaged over the ClinHallu test items of VQA-RAD, PathVQA, MedFrameQA and MedXpertQA (medical VQA with validated reference traces of visual recognition, knowledge recall and reasoning integration); the model writes a structured trace and answer (temperature 0.7) and a Qwen3.5-27B judge grades them (temperature 0.01); lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 2.3 | #164 |
| 2 | Qwen 3.5 9B | 4.8 | #803 |
| 3 | Qwen 3.5 4B | 5.1 | #1175 |
| 4 | Qwen 2.5 VL 7B Instruct | 18.1 | #1833 |
| 5 | MedGemma-4B-IT | 30.5 | #2716 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=clinhallu-reasoning-integration-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-13.