HallusionBench
Multimodal · 2023-10-23
HallusionBench probes language hallucination and visual illusion with paired yes/no questions over figures. Model developers usually report a single headline figure without stating whether it is all-question accuracy or the VLMEvalKit average of question-, figure- and all-question accuracy, so this row is treated as a composite and shares a factor group with the OpenVLM leaderboard protocol.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.5 397B A17B | 71.4 |
| Qwen3.5-27B | 70.0 |
| Kimi K2.5 | 69.8 |
| Qwen3.6 35B A3B | 69.8 |
| Qwen3.5-9B | 69.3 |
| Gemma 4 12B IT | 68.9 |
| Gemini 3 Pro Preview | 68.6 |
| Qwen3.5 35B A3B | 67.9 |
| Qwen3.5 122B A10B | 67.6 |
| Gemma 4 31B IT | 67.4 |
| Qwen3 VL 32B Thinking | 67.4 |
| Qwen3 VL 235B A22B Thinking | 66.7 |
| Qwen3.5-4B | 66.2 |
| Gemma 4 26B A4B IT | 66.1 |
| Qwen3 VL 30B A3B Thinking | 66.0 |
| GPT-5 | 65.7 |
| Gemini 3.5 Flash-Lite | 65.6 |
| LFM3-VL-2B | 65.6 |
| Qwen3 VL 8B Thinking | 65.4 |
| GPT-5.2 | 65.2 |
| Gemini 2.5 Flash-Lite | 64.5 |
| Opus 4.5 | 64.1 |
| Qwen3 VL 4B Thinking | 64.1 |
| Qwen3 VL 32B Instruct | 63.8 |
| Gemini 2.5 Pro | 63.7 |
| GPT-5 Mini | 63.2 |
| Qwen3-VL-235B-A22B-Instruct | 63.2 |
| Qwen3 Omni 30B A3B Thinking | 62.8 |
| Qwen3 VL 30B A3B Instruct | 61.5 |
| Qwen3-VL-8B-Instruct | 61.1 |
| Opus 4.1 | 60.4 |
| Sonnet 4.5 | 59.9 |
| Qwen3 Omni 30B A3B Instruct | 59.7 |
| Sonnet 4 | 59.2 |
| Gemini 2.5 Flash | 59.1 |
| GPT-5.6 Luna | 59.0 |
| GPT-5 Nano | 58.4 |
| Qwen3 VL 4B Instruct | 57.6 |
| InternVL3.5-241B-A28B | 57.3 |
| Gemma 4 E2B IT | 56.7 |