Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding

2026-05-04

AI Summary

Existing long-chain-of-thought (Long-CoT) distillation approaches rely on post-hoc filtering of complete reasoning trajectories, overlooking dynamic collaboration among heterogeneous teacher models and often resulting in sampling redundancy and missed complementary reasoning paths. This work proposes CoRD, a novel framework that, for the first time, introduces a collaborative decoding mechanism at the step level among heterogeneous teachers. By integrating predictive perplexity scoring with beam search, CoRD dynamically guides multiple teachers to jointly generate coherent yet diverse high-quality reasoning trajectories. This approach efficiently constructs structured supervision signals, preserving high-potential hypotheses while reducing data redundancy. Consequently, the student model achieves performance close to that of the teachers with significantly fewer training samples and demonstrates strong generalization capabilities on out-of-domain and open-ended tasks.

Abstract

Distilling large reasoning models is essential for making Long-CoT reasoning practical, as full-scale inference remains computationally prohibitive. Existing curation-based approaches select complete reasoning traces post-hoc, overlooking collaboration among heterogeneous teachers and lacking dynamic exploration, which leads to redundant sampling and missed complementary reasoning. We introduce CoRD, a collaborative multi-teacher decoding framework that performs step-wise reasoning synthesis guided by predictive perplexity-based scoring and beam search. This enables heterogeneous LRMs to jointly construct coherent reasoning trajectories while efficiently preserving diverse, high-potential hypotheses. Experiments show that CoRD produces higher-quality reasoning data and achieves near teacher-level student performance with fewer, structured supervision signals, without substantial efficiency overhead. CoRD further generalizes well to out-of-domain and open-ended settings. The dataset and model are available at \href{https://github.com/DISL-Lab/CoRD}{https://github.com/DISL-Lab/CoRD}.

Problem

• Long-CoT reasoning
• distillation
• multi-teacher collaboration
• reasoning trace curation
• heterogeneous teachers

Innovation

• Long-CoT reasoning
• multi-teacher distillation
• step-wise decoding
• collaborative reasoning synthesis
• perplexity-based scoring