Token-level hallucination detection on MS COCO image captioning (test)
88PrecisionOurs (MLP)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Ours (MLP)LVLM=Qwen2.5-VL-7B2026.04 | 88 | 76 | 81 | 93 | |
| Ours (XGB)LVLM=InternVL2.5-8B2026.04 | 88 | 91 | 90 | 94 | |
| Ours (MLP)LVLM=InternVL2.5-8B2026.04 | 88 | 92 | 90 | 93 | |
| Ours (XGB)LVLM=Qwen2.5-VL-7B2026.04 | 86 | 83 | 84 | 94 | |
| Ours (RF)LVLM=InternVL2.5-8B2026.04 | 85 | 91 | 88 | 93 | |
| ProjectAwayLVLM=InternVL2.5-8B2026.04 | 83 | 81 | 82 | 89 | |
| Ours (XGB)LVLM=LLaVA-1.5-7B2026.04 | 82 | 81 | 82 | 88 | |
| Ours (RF)LVLM=Qwen2.5-VL-7B2026.04 | 82 | 78 | 80 | 93 | |
| ProjectAwayLVLM=LLaVA-1.5-7B2026.04 | 80 | 76 | 79 | 85 | |
| DHCPLVLM=InternVL2.5-8B2026.04 | 80 | 83 | 81 | 87 | |
| Ours (RF)LVLM=LLaVA-1.5-7B2026.04 | 79 | 80 | 80 | 86 | |
| ProjectAwayLVLM=Qwen2.5-VL-7B2026.04 | 78 | 82 | 81 | 89 | |
| DHCPLVLM=LLaVA-1.5-7B2026.04 | 75 | 71 | 73 | 80 | |
| DHCPLVLM=Qwen2.5-VL-7B2026.04 | 75 | 77 | 76 | 86 | |
| SVARLVLM=InternVL2.5-8B2026.04 | 74 | 89 | 80 | 87 | |
| Ours (MLP)LVLM=LLaVA-1.5-7B2026.04 | 72 | 90 | 80 | 86 | |
| HalLocLVLM=Qwen2.5-VL-7B2026.04 | 71 | 73 | 73 | 80 | |
| HalLocLVLM=LLaVA-1.5-7B2026.04 | 70 | 73 | 71 | 76 | |
| MetaToken (GB)LVLM=Qwen2.5-VL-7B2026.04 | 70 | 44 | 54 | 77 | |
| MetaToken (LR)LVLM=InternVL2.5-8B2026.04 | 68 | 62 | 65 | 60 | |
| SVARLVLM=LLaVA-1.5-7B2026.04 | 66 | 89 | 76 | 75 | |
| MetaToken (GB)LVLM=InternVL2.5-8B2026.04 | 66 | 96 | 78 | 72 | |
| MetaToken (GB)LVLM=LLaVA-1.5-7B2026.04 | 65 | 72 | 68 | 72 | |
| HalLocLVLM=InternVL2.5-8B2026.04 | 64 | 68 | 67 | 87 | |
| MetaToken (LR)LVLM=LLaVA-1.5-7B2026.04 | 61 | 43 | 51 | 57 | |
| MetaToken (LR)LVLM=Qwen2.5-VL-7B2026.04 | 46 | 65 | 54 | 69 | |
| SVARLVLM=Qwen2.5-VL-7B2026.04 | 45 | 78 | 57 | 68 |