Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing curation-based approaches to inference are inefficient and fail to adapt dynamically, leading to redundant sampling and missed opportunities for complementary reasoning. |
| Approach: | They propose a collaborative multi-teacher decoding framework that performs step-wise reasoning synthesis guided by predictive perplexity–based scoring and beam search. |
| Outcome: | The proposed framework generates higher-quality reasoning data and achieves student-level results, demonstrating that fine-grained collaboration yields structured, efficient, and robust reasoning distillation. |
Similar Papers
Distilling Reasoning Capabilities into Smaller Language Models (2023.findings-acl)
Copied to clipboard
| Challenge: | a step-by-step reasoning approach like chain of thought has proved to be effective in eliciting reasoning abilities in large language models. |
| Approach: | They propose a knowledge distillation approach that leverages CoT reasoning capabilities of larger models and distills them into smaller models. |
| Outcome: | The proposed scheme boosts the performance of smaller models over 70% on multiple reasoning datasets. |
Mixed Distillation Helps Smaller Language Models Reason Better (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Recent large language models (LLMs) have demonstrated impressive multiple step-by-step reasoning capabilities in recent NLP reasoning tasks. |
| Approach: | They propose a mixed distillation framework that distills multiple step-by-step reasoning abilities into smaller language models (SLMs) they leverage LLMs to generate multiple step by step reasoning rationales by sampling automatically. |
| Outcome: | The proposed framework outperforms existing models on SVAMP, GSM8K and ASDIV, while a single model generated by MD exceeds the comprehensive performance of two individual CoT and PoT distilled models. |
Improving Reasoning Capabilities in Small Models through Mixture-of-layers Distillation with Stepwise Attention on Key Information (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods focus on transferring teacher-generated rationales to student models, but do not explore teachers’ dynamic attention towards critical information during reasoning. |
| Approach: | They propose a method that transfers the teacher’s stepwise attention on key information to the student model and a Mixture of Layers module that allows dynamic alignment between the teacher and student. |
| Outcome: | The proposed framework achieves consistent performance improvements across multiple mathematical and commonsense reasoning datasets. |
DRP: Distilled Reasoning Pruning with Mathematical Skill-aware Step Decomposition for Efficient Large Reasoning Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing solutions to this problem are inference-time pruning and tuning-based distillation. |
| Approach: | They propose a framework that combines inference-time pruning with tuning-based distillation to enable efficient and accurate reasoning. |
| Outcome: | The proposed framework reduces token usage while improving accuracy on GSM8K and AIME tokens while avoiding performance drop. |
Capture the Key in Reasoning to Enhance CoT Distillation Generalization (2025.acl-long)
Copied to clipboard
| Challenge: | Existing distillation methods for Large Language Models (LLMs) focus on fine-tuning student SLMs on correct data, resulting in students struggling to learn the key instead of analyzing mistakes according to correct solutions. |
| Approach: | They propose a method that exposes key reasoning steps rather than simple fine-tuning students' CoTs data by using a set of prompts with similar reasoning paths but divergent conclusions. |
| Outcome: | The proposed method improves student SLMs' ability to learn key reasoning steps rather than fine-tuning them on teacher data. |
MCC-KD: Multi-CoT Consistent Knowledge Distillation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated remarkable abilities in complex reasoning through chain of thought (CoT) prompting. |
| Approach: | They propose to generate multiple rationales for each question and enforce consistency among their predictions by minimizing the bidirectional KL-divergence between the answer distributions. |
| Outcome: | The proposed model achieves superior performance on in-distribution and commonsense reasoning benchmarks. |
PaD: Program-aided Distillation Can Teach Small Models Reasoning Better than Chain-of-thought Fine-tuning (2024.naacl-long)
Copied to clipboard
| Challenge: | Large language models excel in various tasks, but their huge size and inaccessibility of parameters present challenges for practical deployment. |
| Approach: | They propose to use CoT data to distill task-specific ability from large language models to smaller models . they use reasoning programs to suppress errors in distilled data and improve distillation quality . |
| Outcome: | The proposed model outperforms LLMs on arithmetic reasoning, symbolic reasoning, and general ability. |
MoDE-CoTD: Chain-of-Thought Distillation for Complex Reasoning Tasks with Mixture of Decoupled LoRA-Experts (2024.lrec-main)
Copied to clipboard
| Challenge: | Current Chain-of-thought Distillation methods hinder CoT reasoning performance . student models are separately distilled from specific reasoning tasks . parameter update of student models severely harms CoT ability on unseen reasoning tasks. |
| Approach: | They propose a method which distills Chain-of-thought reasoning ability of large language models to much smaller student models. |
| Outcome: | The proposed method improves the reasoning ability of large language models on 14 datasets. |
Improve Student’s Reasoning Generalizability through Cascading Decomposed CoTs Distillation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have tried to distill these capabilities into smaller language models (SLMs) however, these capabilities are often associated with more parameters, which is not practical to emergent in smaller models. |
| Approach: | They propose to decompose the traditional single-step learning process into two cascaded learning steps by restructuring the training objectives and concatenating the question with the rationale as input. |
| Outcome: | Extensive experiments show that the proposed method improves reasoning generalizability and diversity of the model. |
Probe Then Retrieve and Reason: Distilling Probing and Reasoning Capabilities into Smaller Language Models (2024.lrec-main)
Copied to clipboard
| Challenge: | Recent research efforts have focused on distilling Large Language Models into Small Language Model (SLMs) however, the results of CoT distillation are inadequate for knowledge-intensive reasoning tasks. |
| Approach: | They propose a retrieval-based framework which distills question probing and reasoning capabilities from Large Language Models into SLMs. |
| Outcome: | The proposed framework improves probing and reasoning capabilities of large language models in knowledge-intensive reasoning tasks. |