Deep Interaction: An Efficient Human-AI Interaction Method for Correcting LLM Reasoning Errors

Forum topic · 小凯 · 2026-07-17

Summary

Deep Interaction is an efficient human intervention mechanism for precisely correcting reasoning errors in large language models during Chain-of-Thought (CoT) reasoning, proposed by Hefeng Zhou et al. (arXiv 2607.14049). When LLMs make mistakes in multi-step reasoning, current methods either regenerate a response that may err again or require users to laboriously flag faulty steps, often yielding superficial acknowledgments followed by recurring errors. Deep Interaction instead allows direct editing of the original response, correcting erroneous parts while preserving accurate reasoning steps. The edited CoT is refined into a distilled prompt that guides the LLM along the corrected reasoning path. Experiments show over 25% improvement in correction success rate on STEM reasoning tasks and roughly 40% reduction in token usage compared to baseline methods, making human-AI interaction more efficient and reliable.

Paper Overview

Field: Machine Learning Authors: Hefeng Zhou, Jinxuan Zhang, Jiong Lou, Yuxin Liu, Chaochao Lu, Jingjing Qu, Jie Li Published: 2026-07-15 arXiv: 2607.14049

Abstract

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, the authors propose an efficient human intervention mechanism for precisely correcting reasoning errors in LLMs, termed Deep Interaction. The approach enables direct editing of the original response, allowing erroneous parts to be corrected while preserving accurate reasoning steps. The edited CoT is refined into a distilled prompt, which then guides the LLM to proceed along the corrected reasoning path.

Key Results

Over 25% improvement in correction success rate on STEM task reasoning compared to baseline methods
Approximately 40% reduction in token usage