LLM Routing on MMLU Pro Humanities Out-of-Domain
51.74LPMProbeDirichlet
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ProbeDirichletSignal Modality=Hidden-state2026.02 | 51.74 | 64.51 | 11.5 | |
| MaxLogitsSignal Modality=Logit-based2026.02 | 51.19 | 62.15 | 9.25 | |
| EntropySignal Modality=Logit-based2026.02 | 50.77 | 63.01 | 9.23 | |
| SelfAskSignal Modality=Verbose-based2026.02 | 50.58 | 62.94 | 10.75 | |
| EmbeddingMLPSignal Modality=Embedding-based2026.02 | 50.26 | 62.53 | 9.25 | |
| ConfidenceMarginSignal Modality=Logit-based2026.02 | 49.5 | 61.87 | 12.25 | |
| SemanticEntropySignal Modality=Verbose-based2026.02 | 49.42 | 61.64 | 10.75 |