| title | Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding |
|---|---|
| authors | Taewon Yun, Jisu Shin, Jeonghwan Choi, Seunghwan Bang, Hwanjun Song |
| date | 2026-05-04 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2605.02290 |
| affiliation | Data Intelligence System Lab |
| upvotes | 41 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| 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 https://github.com/DISL-Lab/CoRD{https://github.com/DISL-Lab/CoRD}. |
cite
@misc{yun2026distilling,
title={Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding},
author={Taewon Yun and Jisu Shin and Jeonghwan Choi and Seunghwan Bang and Hwanjun Song},
year={2026},
eprint={2605.02290},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2605.02290}
}