ICML 2026 | FusionRoute: From Expert Routing to Self-Correction in Multi-LLM Collaboration
ICML 2026|FusionRoute:从专家路由到自我修正,一种新的多LLM协作范式
FusionRoute proposes a token-level multi-LLM collaboration method that freezes expert models and trains a lightweight router to select an expert for each token while merging router logits with expert logits. The paper evaluates it on GSM8K, MATH-500, HumanEval, MBPP, IfEval, and 500 PerfectBlend prompts.
Why it matters: An ICML 2026 paper proposes FusionRoute, which freezes multiple expert models and trains only a lightweight router that picks experts per token and fuses routing with output scores, a transferable method for multi-model collaboration.