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

TL;DR

The paper presents CoRD, a collaborative multi-teacher decoding method for efficient, high-quality reasoning data distillation in Long-CoT models, outperforming existing approaches.

Contribution

Introducing CoRD, a step-wise, collaborative decoding framework that leverages heterogeneous teachers and dynamic exploration for better reasoning data distillation.

Findings

CoRD produces higher-quality reasoning data.

Achieves near teacher-level student performance with fewer supervision signals.

Generalizes well to 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}.