ACL Findings2026
Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding
Taewon Yun, Jisu Shin, Jeonghwan Choi, Seunghwan Bang, Hwanjun Song
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.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search cites Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
- Small Models Struggle to Learn from Strong Reasoners cites Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
- AlphaZero-Like Tree-Search can Guide Large Language Model Decoding and Training cites Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Builds on9
- Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsJason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma et al.NeurIPS 2022 · 22,562 citations
- AlphaZero-Like Tree-Search can Guide Large Language Model Decoding and TrainingZiyu Wan, Xidong Feng, Muning Wen, Stephen Marcus McAleer et al.ICML 2024 · 325 citations
- Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree SearchHuanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang et al.NeurIPS 2025 · 147 citations
- Small Models Struggle to Learn from Strong ReasonersYuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang et al.ACL Findings 2025 · 5 citations
- Optimizing Test-Time Compute via Meta Reinforcement FinetuningYuxiao Qu, Matthew Y. R. Yang, Amrith Setlur, Lewis Tunstall et al.ICML 2025
Related papers
- Improve Student's Reasoning Generalizability through Cascading Decomposed CoTs DistillationChengwei Dai, Kun Li, Wei Zhou, Songlin HuEMNLP 2024 · 4 citations
- CORD: Bridging the Audio-Text Reasoning Gap via Weighted On-policy Cross-modal DistillationHu Jing, Danxiang Zhu, Xianlong Luo, Dan Zhang et al.ACL Findings 2026
- Think, But Don't Tell: Implicit Reasoning for LLM-based Sequential Recommendation via Multi-Teacher DistillationWeihai Lu, Xiaoxi Cui, Chenke YinSIGIR 2026
- CODI: Compressing Chain-of-Thought into Continuous Space via Self-DistillationZhenyi Shen, Hanqi Yan, Linhai Zhang, Zhanghao Hu et al.EMNLP 2025
- Learning from Diverse Reasoning Paths with Routing and CollaborationZhenyu Lei, Zhen Tan, Song Wang, Yaochen Zhu et al.EMNLP 2025
