Hallucination Diagnosis on HallusionBench
96LI ScoreSmolVLM 500M
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SmolVLM 500M+R=0%2026.05 | 96 | 365 | 278 | 33.85 | — | |
| SmolVLM 500M (Self-Captioning MIT)+R=50%2026.05 | 96 | 254 | 342 | 38.23 | — | |
| SmolVLM 500M (Self-Captioning MIT)+R=25%2026.05 | 94 | 352 | 281 | 33.42 | — | |
| LLaVa-OneVision 4B (Self-Captioning MIT)+R=25%2026.05 | 92 | 185 | 361 | 43.94 | — | |
| LLaVa-OneVision 4B+R=0%2026.05 | 90 | 317 | 273 | 39.45 | — | |
| LLaVa-OneVision 8B (Self-Captioning MIT)+R=25%2026.05 | 89 | 159 | 351 | 46.09 | — | |
| LLaVa-OneVision 8B+R=0%2026.05 | 86 | 218 | 308 | 45.59 | — | |
| LLaVa-OneVision 4B (Self-Captioning MIT)+R=50%2026.05 | 82 | 200 | 349 | 43.99 | — | |
| LLaVa-OneVision 8B (Self-Captioning MIT)+R=50%2026.05 | 82 | 270 | 252 | 45.39 | — | |
| SmolVLM 256M (Self-Captioning MIT)+R=50%2026.05 | 74 | 371 | 278 | 30.94 | — | |
| SmolVLM 2B (Self-Captioning MIT)+R=25%2026.05 | 73 | 210 | 355 | 41.1 | — | |
| SmolVLM 2B (Self-Captioning MIT)+R=50%2026.05 | 71 | 156 | 404 | 42.88 | — | |
| SmolVLM 256M (Self-Captioning MIT)+R=25%2026.05 | 64 | 381 | 281 | 29.43 | — | |
| SmolVLM 2B+R=0%2026.05 | 64 | 349 | 275 | 34.91 | — | |
| SmolVLM 256M+R=0%2026.05 | 55 | 524 | 175 | 26.97 | — | |
| DAPOBackbone=Qwen2.5-VL-3B2026.05 | — | — | — | — | 62.7 | |
| DAPOBackbone=Qwen2.5-VL-7B2026.05 | — | — | — | — | 67 | |
| GRPOBackbone=Qwen2.5-VL-3B2026.05 | — | — | — | — | 62.6 | |
| GRPOBackbone=Qwen2.5-VL-7B2026.05 | — | — | — | — | 66.6 | |
| PACRBackbone=Qwen2.5-VL-3B2026.05 | — | — | — | — | 61.5 | |
| PACRBackbone=Qwen2.5-VL-7B2026.05 | — | — | — | — | 67.6 | |
| PDCRBackbone=Qwen2.5-VL-3B2026.05 | — | — | — | — | 61.8 | |
| PDCRBackbone=Qwen2.5-VL-7B2026.05 | — | — | — | — | 68.4 | |
| Perception-R1Backbone=Perception-R1 (7B)2026.05 | — | — | — | — | 65.4 | |
| Vision-R1Backbone=Vision-R1 (7B)2026.05 | — | — | — | — | 32.2 | |
| Visionary-R1Backbone=Visionary-R1 (3B)2026.05 | — | — | — | — | 30 | |
| Zero-shot InferenceBackbone=Qwen2.5-VL-3B, Mode=Zero-shot2026.05 | — | — | — | — | 45.5 | |
| Zero-shot InferenceBackbone=Qwen2.5-VL-7B, Mode=Zero-shot2026.05 | — | — | — | — | 52.2 |