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

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Hefeng Zhou, Jinxuan Zhang, Jiong Lou, Yuxin Liu, Chaochao Lu, Jingjing Qu, Jie Li
• Venue: arXiv: cs.AI
• Published: 2026-07-15
• Collected: 2026-07-16
• URL: http://arxiv.org/abs/2607.14049v1
• PDF: https://arxiv.org/pdf/2607.14049v1
• DOI: UNKNOWN

Plain-Language Brief

Researchers propose a new method called Deep Interaction to correct errors made by large language models (LLMs) during complex, multi-step tasks. The method allows for direct editing of the original response, improving correction success rate by over 25% and reducing token usage by about 40%.

What the Document Says

• The research introduces Deep Interaction, a method to correct errors in large language models (LLMs).
• Deep Interaction allows for direct editing of the original response.
• The method improves correction success rate by over 25% on STEM tasks reasoning compared to baseline approaches.
• Deep Interaction reduces token usage by approximately 40%.

Key Concepts Mentioned

• Large Language Models (LLMs)
• Chain-of-Thought (CoT) reasoning
• Deep Interaction
• Correction success rate
• Token usage

Claims Made by the Document

• Deep Interaction is an effective method for correcting errors in LLMs.
• Deep Interaction improves correction success rate by over 25%.

Evidence Used

• The text states that Deep Interaction achieves over a 25% improvement in correction success rate.

Methods / Approach

• Deep Interaction method for correcting errors in LLMs involves direct editing of the original response and refining the edited CoT into a distilled prompt.

Limitations Stated by the Authors

The paper does not discuss any limitations of Deep Interaction as presented.

Limitations Noted by the Agent

Agent interpretation — not from the source.

It is unknown if there are potential limitations to the Deep Interaction method that were not discussed by the authors.

Open Questions

• What specific baseline approaches were used for comparison?
• Were any potential drawbacks or challenges associated with implementing Deep Interaction mentioned?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text does not provide information on the specific baseline approaches compared, nor any potential drawbacks or challenges associated with implementing Deep Interaction.

Raw Abstract / Excerpt

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.