LLM Routing on MMLU Pro STEM (Out-of-Domain)
59.2LPMProbeDirichlet
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ProbeDirichletSignal Modality=Hidden-state2026.02 | 59.2 | 71.06 | 15.75 | |
| SemanticEntropySignal Modality=Verbose-based2026.02 | 57.15 | 69.64 | 16 | |
| SelfAskSignal Modality=Verbose-based2026.02 | 57.01 | 69.47 | 13.64 | |
| EmbeddingMLPSignal Modality=Embedding-based2026.02 | 56.78 | 69.29 | 11.42 | |
| ConfidenceMarginSignal Modality=Logit-based2026.02 | 56.64 | 69.47 | 12.5 | |
| MaxLogitsSignal Modality=Logit-based2026.02 | 56.07 | 68.43 | 10.17 | |
| EntropySignal Modality=Logit-based2026.02 | 55.58 | 68.79 | 8.83 |