LLM Routing on Mixed-domain ShareGPT, WildChat, Chatbot Arena 10k episodes (test)
914Average CostAlways-8B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Always-8Beval budget=50002026.05 | 914 | 0.1 | 0 | 2.94 | |
| CQLcost penalty coefficient λ=0.0005, eval budget=50002026.05 | 921 | 0.3 | 0.1 | 2.95 | |
| CQLcost penalty coefficient λ=0.0001, eval budget=50002026.05 | 1,400 | 7.2 | 5.6 | 3.1 | |
| CQLcost penalty coefficient λ=0.00003, eval budget=50002026.05 | 2,528 | 18.1 | 24.6 | 3.32 | |
| CQLcost penalty coefficient λ=0, eval budget=50002026.05 | 3,275 | 24.3 | 39.3 | 3.44 | |
| BCeval budget=50002026.05 | 3,482 | 31.8 | 47.6 | 3.65 | |
| Random (50/50)eval budget=50002026.05 | 3,836 | 35.9 | 52.1 | 3.66 | |
| Heuristiceval budget=50002026.05 | 3,837 | 0.1 | 57.4 | 2.94 | |
| Always-70Beval budget=50002026.05 | 5,146 | 70.9 | 100 | 4.16 |