List your company

Model Routing for Coding Tasks on Agent-as-a-Router In-distribution (test)

57Average Performance (%)Oracle (cost-aware)
35.846441.338246.8352.3218Jun 22, 2026

Evaluation results

MethodLinks
Oracle (cost-aware)Family=Bound2026.065706.958.2
ACRouterFamily=Agentic2026.0649.98205.513.213.79
DimensionBestFamily=Heuristic2026.0647.5277.412.893.69
LogRegFamily=Trained2026.0647.26284.47.546.27
RouteLLM-BERTFamily=Trained2026.0647.22285.57.596.22
kNN RetrievalFamily=Heuristic2026.0647.18286.77.776.07
TF-IDF+MLPFamily=Trained2026.0646.97292.87.696.11
LinUCB (5-seed avg)Family=Bandit2026.0646.84296.910.694.38
LinTS (5-seed avg)Family=Bandit2026.0646.48307.410.354.49
Qwen3.5-0.8B-FinetunedFamily=Trained2026.0646.41309.16.816.82
RouteLLM-MFFamily=Trained2026.0646.16316.57.466.19
Always-Opus 4.6Family=Single-Model2026.0643.83387.134.021.29
RandomFamily=Single-Model2026.0638.75533.615.642.48
Always-Qwen3.5-PlusFamily=Single-Model2026.0637.16580.218.142.05
Always-Kimi-K2.5Family=Single-Model2026.0636.66593.32.912.62