---
title: "Deep Interaction: An Efficient Human-AI Interaction Method for Large Reasoning Models"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.14049v1"
permalink: "https://feed.astralor.com/items/itm_d4d6982a4bb84bd1dc0af415"
published: "2026-07-15T17:16:43.000Z"
score: 75
channel: "ai"
tags: ["CoT","Deep Interaction","Human-Computer Interaction","Large Reasoning Models","Error Correction"]
---

# Deep Interaction: An Efficient Human-AI Interaction Method for Large Reasoning Models

## AI commentary

> This work precisely targets a pain point in current LLM interaction: inefficient error correction. For developers or researchers requiring high-accuracy reasoning, Deep Interaction offers a direct, token-saving correction approach that is especially noteworthy. Although still in preprint, the method is simple and effective, with potential for integration into future tools.

## At a glance

This paper introduces Deep Interaction, a method that allows users to directly edit erroneous steps in the Chain-of-Thought reasoning of large language models while preserving correct parts. The edited CoT is refined into a distilled prompt to steer the model along the corrected path. Experiments on STEM tasks show over 25% improvement in correction success rate and approximately 40% reduction in token usage compared to baselines.

## Key points

- Proposes direct editing of the model's original output, avoiding re-generation or laborious step-by-step flagging.
- Refines the edited CoT into a distilled prompt to guide correct reasoning.
- Achieves >25% improvement in correction success rate and ~40% token reduction on STEM tasks.

## Full text

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 prompt, which then steers the LLM along the corrected reasoning path. Experimental results show that our method achieves over a 25% improvement in correction success rate and reduces token usage by approximately 40% on STEM tasks reasoning compared to baseline approaches.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_d4d6982a4bb84bd1dc0af415