Multiple Choice Question Answering on MMLU-Pro (maj@2, maj@4, maj@8)
46.89Accuracy (maj@2)DLE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DLEBackbone=Qwen2.5-7B-Instruct, Sampler=ε-sampling, Branching Algorithm=DIVFIRST2026.04 | 46.89 | 48.9 | 50.71 | |
| DLEBackbone=Qwen2.5-7B-Instruct, Sampler=ε-sampling, Branching Algorithm=PROBFIRST2026.04 | 46.88 | 48.72 | 50.62 | |
| DLEBackbone=Qwen2.5-7B-Instruct, Sampler=min-p2026.04 | 46.84 | 48.55 | 50.4 | |
| DLEBackbone=Qwen2.5-7B-Instruct, Sampler=top-p+top-k2026.04 | 46.78 | 48.61 | 50.52 | |
| DLEBackbone=Qwen2.5-7B-Instruct, Sampler=ε-sampling, Branching Algorithm=RANDBRANCH2026.04 | 46.7 | 48.72 | 50.69 | |
| Self-consistencyBackbone=Qwen2.5-7B-Instruct, Sampler=ε-sampling2026.04 | 46.19 | 48.31 | 51.89 | |
| Self-consistencyBackbone=Qwen2.5-7B-Instruct, Sampler=min-p2026.04 | 46.14 | 48.09 | 49.88 | |
| Self-consistencyBackbone=Qwen2.5-7B-Instruct, Sampler=top-p+top-k2026.04 | 45.89 | 47.11 | 50.13 | |
| Self-consistencyBackbone=Qwen2.5-7B-Instruct2026.04 | 45.79 | 48.43 | 51.75 |