arXiv 2610.19705
RouteWeaver: Weaving Mode Selection and Execution into Unified LLM Routing
AI17 upvotes
Authors: Xiaohan Wang, Haozhen Zhang, Qingyuan Liu, Tao Feng, Wenya Wang
Overview
The authors propose RouteWeaver, a single LLM routing policy that combines single-round, multi-round, and agentic routing modes and learns both which mode to use and which models to pick within it. They train it with COMET-GRPO, which splits the trajectory-level advantage into separate signals for mode selection and within-mode execution, plus a progressive exploration schedule that moves from externally assigned modes to autonomous choice.
Introduction
Existing LLM routers are typically fixed to one execution paradigm (single-round, multi-round, or agentic), which limits how well they adapt to different queries.
Results
RouteWeaver reaches an average accuracy of 61.4%, an 11.03% relative improvement over the strongest baseline, and beats all baselines in each of five domains. Without progressive exploration, the policy collapses entirely to single-round execution.
Conclusion
Unifying routing modes in one policy with decoupled credit assignment lets the router adapt its execution strategy per query instead of being locked to a single paradigm.