R-Stitch: Dynamic Trajectory Stitching for Efficient Reasoning

TL;DR

R-Stitch is a hybrid decoding framework that dynamically routes reasoning tokens between small and large language models based on entropy, significantly accelerating inference while maintaining accuracy.

Contribution

It introduces a training-free, entropy-guided routing strategy and an adaptive policy extension for efficient, high-quality reasoning in large language models.

Findings

Achieves up to 4.10× speedup with negligible accuracy loss.

Effectively reduces inference complexity by delegating uncertain tokens.

Enables flexible efficiency-accuracy trade-offs without retraining.

Abstract

Chain-of-thought (CoT) enhances the problem-solving ability of large language models (LLMs) but incurs substantial inference cost due to long autoregressive trajectories. Existing acceleration strategies either shorten traces via early stopping or compression, or adopt speculative decoding with a smaller model. However, speculative decoding provides limited gains when model agreement is low and rigidly enforces token-level consistency, overlooking the observation that some smaller models, when correct, produce significantly more concise reasoning traces that could reduce inference length. We introduce R-Stitch, a training-free hybrid decoding framework that leverages token-level entropy as an uncertainty proxy to delegate computation between a small language model (SLM) and an LLM. Our analysis shows that high-entropy tokens are more likely to induce errors, motivating an entropy-guided routing strategy that lets the SLM efficiently handle low-entropy tokens while delegating uncertain ones to the LLM, thereby avoiding full rollbacks and preserving answer quality. We further extend this design with R-Stitch$^{+}$, which learns an adaptive routing policy to adjust the token budget dynamically beyond fixed thresholds. By jointly reducing per-token decoding complexity and the number of generated tokens, our method achieves substantial acceleration with negligible accuracy loss. Concretely, it attains peak speedups of 3.00$\times$ on DeepSeek-R1-Distill-Qwen-7B, 3.85$\times$ on 14B, and 4.10$\times$ on QWQ-32B while maintaining accuracy comparable to full LLM decoding. Moreover, it naturally enables adaptive efficiency--accuracy trade-offs that can be tailored to diverse computational budgets without retraining.