xRouteBench Leaderboard

xRouteBench evaluates routers on both response quality and inference cost under a single protocol, spanning general NLP, memory-augmented, vision, time-series, and personalized scenarios.

Results below are produced by the evaluation pipeline described in Training and evaluation. To reproduce a row, see Evaluation workflow.

Reading the table

• Family groups routers by formulation: rule-based baselines, single-turn, multi-turn, and personalized routers.
• Main track reports the performance-first setting, where the cost weight is zero. Rankings shift substantially once cost enters the objective — no single router wins at every operating point.
• Personalized track is scored on a human-preference split and uses a different router set, so its numbers are not comparable to the main track.

Updating the data

The table is rendered from docs/data/leaderboard.json; no code changes are needed to publish new results. Each row takes a router name, a family, and a scores object keyed by column id. Omit a key to render —, or set "pending": true to grey out a row whose numbers are still in flight.