Knowledge Gap Detection on MMLU
78.2AccuracyAlign-P
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Align-PBackbone=Gemma2-9b2026.04 | 78.2 | 0.534 | |
| GRADEposBackbone=Gemma2-9b2026.04 | 76.8 | 0.838 | |
| Align-PBackbone=Llama3.1-8b2026.04 | 75.9 | 0.506 | |
| Align-PBackbone=Qwen2.5-7b2026.04 | 75.5 | 0.476 | |
| GRADEpreBackbone=Gemma2-9b2026.04 | 75 | 0.809 | |
| GRADEpreBackbone=Qwen2.5-7b2026.04 | 73.5 | 0.688 | |
| GRADEposBackbone=Qwen2.5-7b2026.04 | 73.4 | 0.758 | |
| GRADEposBackbone=Llama3.1-8b2026.04 | 72.7 | 0.778 | |
| P-EntropyBackbone=Qwen2.5-7b2026.04 | 71.3 | 0.678 | |
| GRADEpreBackbone=Llama3.1-8b2026.04 | 70.3 | 0.751 | |
| P-EntropyBackbone=Gemma2-9b2026.04 | 70.2 | 0.765 | |
| P-EntropyBackbone=Llama3.1-8b2026.04 | 70.1 | 0.728 | |
| JudgeBackbone=Llama3.1-8b2026.04 | 51.9 | — | |
| ICBackbone=Gemma2-9b2026.04 | 50.5 | 0.567 | |
| ICBackbone=Llama3.1-8b2026.04 | 50.4 | 0.651 | |
| ICBackbone=Qwen2.5-7b2026.04 | 50.2 | 0.536 | |
| JudgeBackbone=Qwen2.5-7b2026.04 | 49 | — | |
| JudgeBackbone=Gemma2-9b2026.04 | 47.2 | — |