LLM Routing on Magpie Out-of-Domain (LPM, MPM, HCR)
63.53LPM ScoreProbeDirichlet
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| ProbeDirichletSignal Modality=Hidden-state2026.02 | 63.53 | 71.77 | 14.5 | |
| SelfAskSignal Modality=Verbose-based2026.02 | 63.35 | 71.91 | 13.5 | |
| MaxLogitsSignal Modality=Logit-based2026.02 | 63.13 | 71.63 | 11 | |
| SemanticEntropySignal Modality=Verbose-based2026.02 | 63.08 | 70.84 | 14.5 | |
| EntropySignal Modality=Logit-based2026.02 | 63.08 | 71.65 | 11 | |
| EmbeddingMLPSignal Modality=Embedding-based2026.02 | 62.66 | 71.63 | 9 | |
| ConfidenceMarginSignal Modality=Logit-based2026.02 | 62.6 | 71.34 | 9.5 |