[论文] Deep Interaction: An Efficient Human-AI Interaction Method for Large R...

论文概要

研究领域: ML
作者: Hefeng Zhou, Jinxuan Zhang, Jiong Lou, Yuxin Liu, Chaochao Lu, Jingjing Qu, Jie Li
发布时间: 2026-07-15
arXiv: 2607.14049

中文摘要

思维链（CoT）推理的出现显著增强了大语言模型（LLM）处理复杂多步任务的能力。然而，当错误发生时，当前交互方法通常涉及重新生成另一个可能再次出错的响应，或用户在后续轮次中费力标记错误步骤，可能得到<你是对的，我在这里犯了错误>的响应后类似错误再次出现。为解决此问题，我们提出一种高效的人为干预机制，用于精确纠正LLM中的推理错误，称为Deep Interaction。我们的方法允许直接编辑原始响应，允许纠正错误部分同时保留准确推理步骤。我们将编辑后的CoT提炼为蒸馏提示，然后引导LLM沿纠正的推理路径前进。实验结果表明，与基线方法相比，我们的方法在STEM任务推理中实现超过25%的纠正成功率提升，并将token使用量减少约40%。

原文摘要

The emergence of Chain-of-Thought (CoT) reasoning has significantly enhanced the ability of large language models (LLMs) to tackle complex, multi-step tasks. However, when errors occur, current interaction approaches typically involve re-generating another response that may make mistakes again, or users laboriously flag the faulty step in follow-up turns that may get responses <You are right, I made a mistake here> followed by similar errors recurring. To address this issue, we propose an efficient human intervention mechanism for precisely correcting reasoning errors in LLMs, termed Deep Interaction. Our approach enables direct editing of the original response, allowing erroneous parts to be corrected while preserving accurate reasoning steps. We refine the edited CoT into a distilled pro...

自动采集于 2026-07-17

#论文 #arXiv #ML #小凯