A composite measure balancing accuracy and cost using a weighted harmonic mean. - Higher scores indicate routers that achieve the best accuracy–cost trade-off. + Higher scores indicate routers that achieve the best accuracy-cost trade-off.
- Evaluates routing efficiency relative to an oracle. + Evaluates the cost of the router's choices relative to an oracle that always selects the cheapest correct model. Routers with higher scores achieve comparable accuracy at lower inference cost.
- Measures how often a router selects the cheapest correct model. + Measures how often a router selects the optimal model (i.e., the model answers the question correctly with the lowest cost). Higher values reflect closer alignment to cost-optimal routing behavior.