LLM Routing on MATH Out-of-Domain
73.9AUROCProbeDirichlet
Evaluation Results
| Method | Links | |
|---|---|---|
| ProbeDirichletsignal_modality=Hidden-state2026.02 | 73.9 | |
| EmbeddingMLPsignal_modality=Embedding-based2026.02 | 56.97 | |
| Entropysignal_modality=Logit-based2026.02 | 55.3 | |
| SemanticEntropysignal_modality=Verbose-based2026.02 | 55.25 | |
| ConfidenceMarginsignal_modality=Logit-based2026.02 | 50.05 | |
| SelfAsksignal_modality=Verbose-based2026.02 | 49.29 | |
| MaxLogitssignal_modality=Logit-based2026.02 | 47 |