R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing

TL;DR

R2R introduces a token routing method that selectively employs large models only for divergent reasoning tokens, significantly improving efficiency while maintaining high accuracy in complex tasks.

Contribution

The paper presents R2R, a neural token routing approach that efficiently combines small and large language models by identifying and focusing on divergent reasoning tokens.

Findings

R2R surpasses R1-7B accuracy by 1.6x with fewer parameters.

Achieves 2.8x speedup over R1-32B with similar performance.

Most tokens are neutral or identical, enabling efficient routing.

Abstract

Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token routing method that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6x, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8x wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency. Our code is available at https://github.com/thu-nics/R2R.