Token-level Hallucination Detection on AIME 2025
6.46S_incorQwen3-0.6B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Qwen3-0.6BCategory=Backbone Models2026.05 | 6.46 | 95.84 | |
| Qwen3-1.7BCategory=Backbone Models2026.05 | 18.76 | 100 | |
| R1-Qwen3-8BCategory=Critic Models2026.05 | 40.14 | 100 | |
| Qwen3-4BCategory=Backbone Models2026.05 | 44.75 | 100 | |
| Qwen3-8BCategory=Backbone Models2026.05 | 48.92 | 99.72 | |
| QwQ-32BCategory=Critic Models2026.05 | 56.18 | 99.31 | |
| GPT-4.1Category=Critic Models2026.05 | 56.24 | 100 | |
| TOKENHD-0.6BCategory=Detector Models2026.05 | 64.34 | 89.31 | |
| TOKENHD-1.7BCategory=Detector Models2026.05 | 67.02 | 96.41 | |
| TOKENHD-4BCategory=Detector Models2026.05 | 67.32 | 92.66 | |
| TOKENHD-8BCategory=Detector Models2026.05 | 68.88 | 95.84 | |
| o4-miniCategory=Critic Models2026.05 | 76.09 | 100 | |
| o3Category=Labeler Models2026.05 | 85.88 | 100 | |
| GPT-5Category=Labeler Models2026.05 | 87.89 | 100 |