LLM Routing on MMLU In-Domain (LPM, MPM, HCR)
78.51LPMProbeDirichlet
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ProbeDirichletSignal Modality=Hidden-state2026.02 | 78.51 | 84.31 | 21 | |
| SelfAskSignal Modality=Verbose-based2026.02 | 77.52 | 83.92 | 12.5 | |
| SemanticEntropySignal Modality=Verbose-based2026.02 | 75.9 | 82.44 | 15.5 | |
| EmbeddingMLPSignal Modality=Embedding-based2026.02 | 75.9 | 82.61 | 10 | |
| ConfidenceMarginSignal Modality=Logit-based2026.02 | 75.7 | 82.34 | 10 | |
| EntropySignal Modality=Logit-based2026.02 | 75.29 | 82.25 | 9 | |
| MaxLogitsSignal Modality=Logit-based2026.02 | 75.03 | 81.61 | 8 |