Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866918480180150272 |
|---|---|
| author | Yun, Taewon Shin, Jisu Choi, Jeonghwan Bang, Seunghwan Song, Hwanjun |
| author_facet | Yun, Taewon Shin, Jisu Choi, Jeonghwan Bang, Seunghwan Song, Hwanjun |
| contents | 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}. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_02290 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding Yun, Taewon Shin, Jisu Choi, Jeonghwan Bang, Seunghwan Song, Hwanjun Artificial Intelligence 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}. |
| title | Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2605.02290 |