CriticRouter: Criticality-aware Step Router for Small-Large Collaborative Reasoning

Abstract

Large reasoning models (LRMs) achieve strong performance on complex reasoning tasks, but their long chain-of-thought trajectories incur substantial inference latency and computational cost. Small-large model collaborative reasoning mitigates this issue by offloading part of the reasoning steps to a small language model (SLM) while preserving large language model (LLM) computation for more demanding steps. Existing draft-and-verify methods use an external verifier to reject low-scoring SLM steps and regenerate them with the LLM, introducing per-step verification overhead. Moreover, we observe that many LLM rewrites remain semantically close to the rejected SLM drafts, causing unnecessary decoding cost. Further analysis shows that these redundant rewrites often occur on local derivation steps, whose content is largely constrained by recent context. Motivated by these observations, we propose CriticRouter, a verifier-free step-wise routing framework that reduces redundancy based on structural criticality as a routing signal. Specifically, structural criticality measures how sensitive the SLM's next-step generation is to perturbations in earlier reasoning history, and low-criticality steps are viewed as locally stable and delegated to the SLM, whereas high-criticality steps are treated as history-sensitive and routed to the LLM. To reduce probing overhead, CriticRouter further adopts short concurrent probes and progressive mask advancement for better cache reuse. Experiments on multiple benchmarks show that CriticRouter achieves comparable accuracy to standard LLM reasoning while significantly reducing latency, demonstrating a better utility-efficiency trade-off.