reasoning benchmark
Hallusion Bench
A comprehensive benchmark designed to evaluate image-context reasoning in large visual-language models (LVLMs) by challenging models with 346 images and 1,129 carefully crafted questions to assess language hallucination and visual illusion
Updated Aug 7, 2026
Published models16
Registry coverage16
MetricScore
EvidenceB
Hallusion Bench leaderboard
Sorted by the source-provided rank. Higher score is better according to the registry.
| 1 | AC | 70.0% | 100.0% | 16 | C | |
| 2 | AC | 69.8% | 93.3% | 16 | C | |
| 3 | AC | 67.9% | 86.7% | 16 | C | |
| 4 | AC | 67.6% | 80.0% | 16 | C | |
| 5 | AC | 67.4% | 73.3% | 16 | C | |
| 6 | AC | 66.7% | 66.7% | 16 | C | |
| 7 | AC | 66.0% | 60.0% | 16 | C | |
| 8 | AC | 65.4% | 53.3% | 16 | C | |
| 9 | AC | 64.1% | 46.7% | 16 | C | |
| 10 | AC | 63.8% | 40.0% | 16 | C | |
| 11 | AC | 63.2% | 33.3% | 16 | C | |
| 12 | AC | 61.5% | 26.7% | 16 | C | |
| 13 | AC | 61.1% | 20.0% | 16 | C | |
| 14 | AC | 57.6% | 13.3% | 16 | C | |
| 15 | AC | 55.2% | 6.7% | 16 | C | |
| 16 | AC | 52.9% | 0.0% | 16 | C |
Score distribution
Top published rows on the benchmark's original scale.
Hallusion Bench highlights
The top published results on this benchmark's own scale.
What is Hallusion Bench?
Definition and scoring fields from the benchmark registry.
A comprehensive benchmark designed to evaluate image-context reasoning in large visual-language models (LVLMs) by challenging models with 346 images and 1,129 carefully crafted questions to assess language hallucination and visual illusion
Scores are shown in ratio. The current registry marks this benchmark as not independently verified with evidence level B.
- Family
- Hallusion Bench
- Modality
- multimodal
- Primary category
- reasoning
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- hallusion-bench|llm-stats-current
Source-native results are preserved. Eligibility for the overall LLMBoard score is a separate policy decision.
FAQ
Common questions about Hallusion Bench.
Which model scores highest on Hallusion Bench?
Qwen3.5-27B is currently ranked first with 70.0%.
What does Hallusion Bench measure?
A comprehensive benchmark designed to evaluate image-context reasoning in large visual-language models (LVLMs) by challenging models with 346 images and 1,129 carefully crafted questions to assess language hallucination and visual illusion
Is a higher score better?
Yes. Higher values rank better for this benchmark.
How many models are compared?
16 unique published model results are currently shown.
Does this benchmark affect the overall score?
This benchmark is preserved as source-native evidence but is not eligible for the current overall score.