Hallucination Evaluation on POPE adversarial (400-sample held-out set)
89.3AccuracyLLaVA-v1.6
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| LLaVA-v1.6Config=k = 3, s=0.52026.05 | 89.3 | 6.5 | 0.8 | 1.5 | — | |
| LLaVA-v1.6Config=k = 5, s=0.52026.05 | 89 | 6.5 | 0.5 | 1.5 | — | |
| LLaVA-v1.6Config=k = 8, s=0.52026.05 | 89 | 3 | 0.5 | 2 | 40 | |
| LLaVA-v1.6Config=k = 15, s=0.52026.05 | 88.8 | 4 | 0.3 | 1 | 20 | |
| LLaVA-v1.6Config=Baseline2026.05 | 88.5 | 5 | — | — | — | |
| LLaVA-v1.6Config=s=0.5, top-10, Intervention=Targeted suppression2026.05 | 88.5 | 3 | 0 | 2 | 40 | |
| InternVL3-8BConfig=Baseline2026.05 | 88.4 | 15.6 | — | — | — | |
| Qwen3-VLConfig=Baseline2026.05 | 88.2 | 8.4 | — | — | — | |
| InternVL3-14BConfig=Baseline2026.05 | 88 | 15.5 | — | — | — | |
| InternVL3-14BConfig=s=0.5, all 16, Intervention=Targeted suppression2026.05 | 87.8 | 9 | 0.3 | 6.5 | 42 | |
| Qwen3-VLConfig=s=0.25, all 12, Intervention=Targeted suppression2026.05 | 87.2 | 4 | 1 | 4.4 | 52 | |
| InternVL3-8BConfig=s=0.5, all 6, Intervention=Targeted suppression2026.05 | 86.4 | 5.6 | 2 | 10 | 64 | |
| Llama-3.2Config=Baseline2026.05 | 84.8 | 18.4 | — | — | — | |
| Llama-3.2Config=s=0.5, all 20, Intervention=Targeted suppression2026.05 | 84.6 | 4.4 | 0.2 | 14 | 76 | |
| LLaVA-v1.6Config=k = 20, s=0.52026.05 | 84 | 3.5 | 4.5 | 1.5 | 30 | |
| LLaVA-v1.6Config=k = 30 (all), s=0.52026.05 | 79.5 | 1.5 | 9 | 3.5 | 70 |