RouteWeaver trains one LLM policy to pick routing modes

What the research found

RouteWeaver is an arXiv research paper proposing a single LLM policy that jointly picks among single-round, multi-round, and agentic routing modes, trained with COMET-GRPO and progressive exploration; it reports 61.4% average accuracy, 11.03% relative gain over the strongest baseline.

Claims and how well the evidence supports them

1. RouteWeaver achieves an average accuracy of 61.4% across mathematics, coding, knowledge, logical reasoning, and factual recall 99%
2. RouteWeaver is an 11.03% relative improvement over the strongest baseline 98%
3. Without progressive exploration the routing policy collapses to single-round execution 99%
4. RouteWeaver integrates single-round, multi-round, and agentic routing modes in one LLM policy 97%

Evidence

• huggingface.co https://huggingface.co/papers/2610.19705
• arxiv.org https://arxiv.org/abs/2610.19705

Actionability rated low.