LLM Routing on MATH Out-of-Domain (LPM, MPM, HCR)
69.24LPM ScoreProbeDirichlet
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ProbeDirichletSignal Modality=Hidden-state2026.02 | 69.24 | 81.45 | 21 | |
| EmbeddingMLPSignal Modality=Embedding-based2026.02 | 63.95 | 78.18 | 13.5 | |
| EntropySignal Modality=Logit-based2026.02 | 63.81 | 77.84 | 12.5 | |
| ConfidenceMarginSignal Modality=Logit-based2026.02 | 62.72 | 77.26 | 10 | |
| SemanticEntropySignal Modality=Verbose-based2026.02 | 61.63 | 76.24 | 13 | |
| SelfAskSignal Modality=Verbose-based2026.02 | 61.46 | 75.34 | 11.5 | |
| MaxLogitsSignal Modality=Logit-based2026.02 | 61.16 | 76.63 | 10 |