arXiv 2601.035592026-04-19cs.CL

DiffCoT：在大语言模型中采用扩散风格的推理链推理

DiffCoT: Diffusion-styled Chain-of-Thought Reasoning in LLMs

Hong Kong Baptist University（香港 Baptist 大学）；Harbin Institute of Technology（哈尔滨工业大学）

Shidong Cao, Hongzhan Lin, Yuxuan Gu, Ziyang Luo, Jing Ma

备注：: DiffCoT improves multi-step LLM reasoning by applying diffusion-based iterative denoising to correct intermediate Chain-of-Thought steps
发表信息：: The 64th Annual Meeting of the Association for Computational Linguistics 2026

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

DiffCoT通过扩散风格的推理链框架，改进了大语言模型在多步数学问题解决中的鲁棒性和错误纠正能力，通过迭代去噪过程实现中间步骤的统一生成与回顾性修正。

摘要：

Chain-of-Thought (CoT) 推理提高了大语言模型在多步数学问题解决中的能力，但仍然容易受到暴露偏差和误差累积的影响，因为早期的错误会不可逆地传播通过自回归解码。在本工作中，我们提出了DiffCoT，一种扩散风格的CoT框架，将Co T推理重新表述为迭代去噪过程。DiffCoT通过滑动窗口机制在推理步骤层面整合扩散原理，使中间步骤的统一生成和回顾性修正成为可能，同时保持令牌级别的自回归性。为了保持因果一致性，我们进一步引入了一个因果扩散噪声调度，该调度尊重推理链的时间结构。在三个跨多样模型背骨的多步CoT推理基准上的广泛实验表明，DiffCoT在现有CoT偏好优化方法上始终表现更优，实现了在CoT推理中的改进鲁棒性和错误纠正能力。

英文摘要：

Chain-of-Thought (CoT) reasoning improves multi-step mathematical problem solving in large language models but remains vulnerable to exposure bias and error accumulation, as early mistakes propagate irreversibly through autoregressive decoding. In this work, we propose DiffCoT, a diffusion-styled CoT framework that reformulates CoT reasoning as an iterative denoising process. DiffCoT integrates diffusion principles at the reasoning-step level via a sliding-window mechanism, enabling unified generation and retrospective correction of intermediate steps while preserving token-level autoregression. To maintain causal consistency, we further introduce a causal diffusion noise schedule that respects the temporal structure of reasoning chains. Extensive experiments on three multi-step CoT reasoning benchmarks across diverse model backbones demonstrate that DiffCoT consistently outperforms existing CoT preference optimization methods, yielding improved robustness and error-correction capability in CoT reasoning.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。