RouteWeaver jointly learns LLM routing mode and execution
Original title: RouteWeaver: Weaving Mode Selection and Execution into Unified LLM Routing
Average accuracy is 61.4%, an 11.03% relative gain over the strongest baseline across five domains.
RouteWeaver puts single-round, multi-round, and agentic routing inside one LLM policy that chooses both the mode and the models used in that mode. COMET-GRPO splits trajectory-level advantage into separate signals for mode selection and within-mode execution. Progressive exploration starts with assigned modes and later lets the router choose them. Without progressive exploration, the policy collapses to single-round execution.
- Evaluation covers mathematics, coding, knowledge, logical reasoning, and factual recall.
- RouteWeaver beats every compared baseline in each of the five domains.
- Average accuracy is 61.4%, 11.03% relatively above the strongest baseline.
- Ablations say both COMET-GRPO and progressive exploration matter for adaptive routing.
From the original article
1 min read at huggingface.co
LLM routing dynamically selects and coordinates heterogeneous language models across diverse tasks. However, existing approaches are largely built around a predefined execution paradigm, such as single-round, multi-round, or agentic routing, which fixes how model calls are organized and limits adaptation to different query needs.
By Xiaohan Wang, Haozhen Zhang, Qingyuan Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.