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

2026-07-15

AI Summary

Current large language models struggle to efficiently correct erroneous reasoning steps, often repeating mistakes and incurring high interaction costs. This work proposes a Deep Interaction mechanism that, for the first time, enables users to perform precise local edits on the model’s original chain of thought. The corrected reasoning chain is then distilled into a guiding prompt that steers the model to re-reason along the accurate path. Built upon the Chain-of-Thought framework, the approach integrates response editing, prompt distillation, and guided re-reasoning. Evaluated on STEM tasks, it achieves a greater than 25% improvement in error-correction success rate while reducing token consumption by approximately 40%, substantially enhancing both the efficiency and accuracy of error correction.

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, 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.

Problem

• Chain-of-Thought
• reasoning errors
• human-AI interaction
• large language models
• error correction

Innovation

• Deep Interaction
• Chain-of-Thought
• Human-AI Interaction
• Reasoning Correction
• Prompt Distillation