Agentic Model Routing on In-Distribution n=2,919 (test)
57Average Performance (%)Oracle
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| OracleTaxonomy=Oracle2026.06 | 57 | 0 | 8.2 | |
| ACRouterTaxonomy=Agent-as-a-Router2026.06 | 49.98 | 205.5 | 3.79 | |
| DimensionBestTaxonomy=Static: Heuristic2026.06 | 47.5 | 277.4 | 3.69 | |
| LogRegTaxonomy=Static: Trained Policy2026.06 | 47.26 | 284.4 | 6.27 | |
| RouteLLM-BERTTaxonomy=Static: Trained Policy2026.06 | 47.22 | 285.5 | 6.22 | |
| kNN RetrievalTaxonomy=Static: Heuristic2026.06 | 47.18 | 286.7 | 6.07 | |
| TF-IDF+MLPTaxonomy=Static: Trained Policy2026.06 | 46.97 | 292.8 | 6.11 | |
| LinUCBTaxonomy=Dynamic: Online Bandit2026.06 | 46.84 | 296.9 | 4.38 | |
| LinTSTaxonomy=Dynamic: Online Bandit2026.06 | 46.48 | 307.4 | 4.49 | |
| Qwen3.5-0.8B-FinetunedTaxonomy=Static: Trained Policy2026.06 | 46.41 | 309.1 | 6.82 | |
| RouteLLM-MFTaxonomy=Static: Trained Policy2026.06 | 46.16 | 316.5 | 6.19 | |
| Always-Opus 4.6Taxonomy=Single-Model Baselines2026.06 | 43.83 | 387.1 | 1.29 | |
| RandomTaxonomy=Single-Model Baselines2026.06 | 38.75 | 533.6 | 2.48 | |
| Always-Qwen3.5-PlusTaxonomy=Single-Model Baselines2026.06 | 37.16 | 580.2 | 2.05 | |
| Always-Kimi-K2.5Taxonomy=Single-Model Baselines2026.06 | 36.66 | 593.3 | 12.62 |