Performance Change Analysis on GSM8K, MATH-500, MMLU, and CNN/DM
4.01Mean DeltaTop-k
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Top-kBackbone=Llama-3.1-8B/Llama-3.2-1B, k=22025.09 | 4.01 | -10.7 | |
| SelfJudge-FBackbone=Llama-3.1-8B/Llama-3.2-1B, Thresholding strategy=best F1 score2025.09 | 2.17 | -1.6 | |
| AutoJudge-FBackbone=Llama-3.1-8B/Llama-3.2-1B, Thresholding strategy=best F1 score2025.09 | 1.7 | -2.3 | |
| SelfJudge-RBackbone=Llama-3.1-8B/Llama-3.2-1B, Thresholding strategy=best recall2025.09 | 0.81 | -0.5 | |
| AutoJudge-RBackbone=Llama-3.1-8B/Llama-3.2-1B, Thresholding strategy=best recall2025.09 | 0.49 | -1.4 |