将大模型推理过程建模为细粒度有向无环图，揭示其内在对话结构。
ReasoningFlow: Discourse Structures for Understanding LLM Reasoning Traces
- 将非线性推理轨迹转化为细粒度有向无环图，捕捉回溯与自修正等结构。
- 分析发现错误步骤大多未用于最终答案，且不同模型推理结构高度相似。
- 可识别局部验证、自我反思等行为，提升推理过程监控能力。
大型推理模型（LRMs）的推理轨迹具有非线性结构，如回溯和自修正，增加了评估与监控的难度。我们提出 ReasoningFlow 框架，将 LRM 推理轨迹中的对话结构建模为细粒度有向无环图（DAG）。通过人工标注 31 条轨迹（共 2.1 千步），达成高一致性的标注协议，并扩展至自动标注 1,260 条轨迹（247.7 千步），涵盖数学、科学与论辩三类任务，以及 Qwen2.5-32B-Inst、QwQ-32B、DeepSeek-V3、DeepSeek-R1、GPT-oss-120B 五种模型。分析显示：(1) 尽管训练基础模型不同且后训练数据可能无重叠，但各类模型的推理结构仍具高度相似性；(2) ReasoningFlow 可揭示多样化的细粒度推理行为（如局部验证、自我反思、假设生成），提升推理轨迹可监控性；(3) 多数错误步骤未被用于推导最终答案；(4) 步骤间的机制因果关系并不反映语言层面的对话结构。代码与数据集已公开于 https://github.com/jinulee-v/reasoningflow。
原文摘要 · Abstract (English)
Large reasoning models (LRMs) produce reasoning traces with non-linear structures, such as backtracking and self-correction, that complicate the evaluation and monitoring of the reasoning process. We introduce ReasoningFlow, a framework that captures the discourse structures of LRM reasoning traces into fine-grained directed acyclic graphs (DAGs). We develop and validate our annotation schema through careful manual annotation of 31 traces (2.1k steps), achieving high inter-annotator agreement, then scale to automatic annotation of 1,260 traces (247.7k steps) spanning three tasks (math, science, argumentation) and five models (Qwen2.5-32B-Inst, QwQ-32B, DeepSeek-V3, DeepSeek-R1, GPT-oss-120B). By analyzing ReasoningFlow graphs, we find: (1) LRMs exhibit structurally similar traces, despite being trained from different base models and potentially non-overlapping post-training data. (2) ReasoningFlow reveals diverse fine-grained reasoning behaviors (e.g., local verification, self-reflection, and assumptions) that can be used for better reasoning trace monitorability. (3) In LRMs, most of the erroneous steps are not used to derive final answers. (4) Mechanistic causal dependencies between steps do not reflect the language-level discourse structure. We release the dataset and code in: https://github.com/jinulee-v/reasoningflow.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。