Research map / Multi-agent systems and games / LLM agents
MACoT: Synthesizing Chains of Thought for Small Models via Multi-Agent Collaboration
Guokai Tang, Feng Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MACoT, a multi-agent collaboration framework that synthesizes chains of thought for small language models as an alternative to distilling long CoTs from large reasoning models.
Problem
Distilling long chains of thought from large reasoning models into small language models is common but the abstract cuts off before stating the specific shortcomings; only the motivation (SLMs are efficient but need better reasoning) is stated.
Approach
Synthesizes chain-of-thought training data for small models through collaboration among multiple agents rather than copying long-CoTs from large reasoning models. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Problem Passport analysis stage (ablation) | MACoT · AIME | Pass@1 accuracy (%): 8.3 -> 11.67 | Improved |
| add the Problem Passport analysis stage (ablation) | MACoT · AMC23 | Pass@1 accuracy (%): 45.0 -> 52.5 | Improved |
| add one self-reflection paragraph via Student-Teacher refinement (ablation) | MACoT · BBH | Pass@1 accuracy (%): 61.5 -> 63.64 | Improved |
| add one self-reflection paragraph via Student-Teacher refinement (ablation) | MACoT · MATH500 | Pass@1 accuracy (%): 76.6 -> 77.4 | Improved |
| increase self-reflection paragraphs beyond one (N2, N4) (ablation) | MACoT · AIME24 and MATH-500 | Pass@1 accuracy (%): N2 average drop 1.8 points, N4 average drop 2.4 points | Hurt |
| synthesize CoTs with MACoT instead of using original LIMO CoTs (same 817 problems) | Qwen2.5-7B-Instruct · BBH | Pass@1 accuracy (%): +2.68 points | Improved |
| synthesize CoTs with MACoT instead of using original LIMO CoTs (same 817 problems) | Qwen2.5-7B-Instruct · GPQA and Omni-Math | Pass@1 accuracy (%) | Hurt |
| fine-tune on MACoT CoTs instead of R1 long-CoTs (same 1879 problems) | Qwen2.5-7B-Instruct · Average of 9 benchmarks | Avg. accuracy (%): 42.2 -> 50.04 | Improved |
| fine-tune on MACoT CoTs instead of R1 long-CoTs | Qwen2.5-7B-Instruct · Average of 9 benchmarks | Avg. tokens: 4972 -> 745 | Improved |
| fine-tune on MACoT versus base model | Qwen2.5-7B-Instruct · Gaokao | Pass@1 accuracy (%): 41.31 -> 41.31 | No clear effect |
| fine-tune on Numina CoTs versus base model | Qwen2.5-7B-Instruct · Average of 9 benchmarks | Avg. accuracy (%): 46.8 -> 39.7 | Hurt |
| fine-tune on R1 long-CoT distillation versus base model | Qwen2.5-7B-Instruct · Average of 9 benchmarks | Avg. accuracy (%): 46.8 -> 42.2 | Hurt |
Closest papers by meaning
- Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.91)
- SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.91)
- Mind the Gap: Bridging Thought Leap for Improved Chain-of-Thought Tuning (NeurIPS 2025 · similarity 0.90)
- Compositional Generalization from Learned Skills via CoT Training: A Theoretical and Structural Analysis for Reasoning (ICLR 2026 · similarity 0.89)
- Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.89)
- Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
