LLM Routing on MMLU Pro Out-of-Domain
65.32STEM ScoreProbeDirichlet
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| ProbeDirichletsignal_modality=Hidden-state2026.02 | 65.32 | 57.84 | 58.82 | 62.77 | |
| SemanticEntropysignal_modality=Verbose-based2026.02 | 56.27 | 51.72 | 52.9 | 53.95 | |
| ConfidenceMarginsignal_modality=Logit-based2026.02 | 54.42 | 46.97 | 54.37 | 49.52 | |
| SelfAsksignal_modality=Verbose-based2026.02 | 53.74 | 55.86 | 56.06 | 50.91 | |
| EmbeddingMLPsignal_modality=Embedding-based2026.02 | 52.97 | 53.77 | 48.16 | 50.45 | |
| MaxLogitssignal_modality=Logit-based2026.02 | 50.03 | 50.53 | 41.14 | 46.43 | |
| Entropysignal_modality=Logit-based2026.02 | 49.7 | 52.36 | 48.54 | 49.23 |