LLM Routing on MMLU Pro Social Sciences (Out-of-Domain)
59.2LPMSelfAsk
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SelfAskSignal Modality=Verbose-based2026.02 | 59.2 | 69.66 | 11 | |
| ProbeDirichletSignal Modality=Hidden-state2026.02 | 59.12 | 69.16 | 14.83 | |
| ConfidenceMarginSignal Modality=Logit-based2026.02 | 58.81 | 68.36 | 11.68 | |
| SemanticEntropySignal Modality=Verbose-based2026.02 | 57.4 | 67.71 | 13.33 | |
| EntropySignal Modality=Logit-based2026.02 | 57.1 | 67.69 | 10.5 | |
| EmbeddingMLPSignal Modality=Embedding-based2026.02 | 56.38 | 67.15 | 9.67 | |
| MaxLogitsSignal Modality=Logit-based2026.02 | 55.24 | 66.13 | 7.5 |