Deep Interaction proposes direct editing for correcting LLM reasoning

Introduction

Chain-of-Thought reasoning has become a central technique for making large language models handle multi-step tasks, from math problems to scientific reasoning. Yet the user experience remains fragile when a model makes a mistake halfway through its reasoning. The usual options are limited: ask the model to regenerate the whole answer, or explain in a follow-up message where the error occurred. Both approaches can be inefficient. Regeneration may discard correct work, while conversational correction can still lead the model back into a similar mistake.

The arXiv paper “Deep Interaction: An Efficient Human-AI Interaction Method for Large Reasoning Models” proposes a different interaction pattern. Instead of treating correction as another turn in a chat, Deep Interaction allows users to directly edit the model’s original reasoning trace. The system then turns the edited Chain-of-Thought into a distilled prompt that guides the model along the corrected reasoning path.

Key points

• Local repair instead of full regeneration: The method focuses on keeping the parts of the reasoning that are already correct and intervening only where the error occurs. This is a more targeted alternative to asking the model to start over.
• Direct editing of the original response: Users do not need to describe the mistake indirectly in a new message. They can modify the faulty reasoning segment in place, making the correction signal more explicit.
• Distilling the edited CoT into a prompt: Deep Interaction does not merely append the full edited reasoning trace to the context. According to the paper, it refines the edited CoT into a distilled prompt that steers the model’s next reasoning steps.
• Evaluation on STEM reasoning: The authors report that, on STEM task reasoning, the approach improves correction success rate by more than 25% and reduces token usage by approximately 40% compared with baseline approaches.

Why it matters

The significance of Deep Interaction is less about adding another reasoning trick and more about redesigning how humans correct reasoning models. In complex tasks, a wrong final answer often originates from a single flawed intermediate step: a misread condition, an invalid transformation, or a mistaken assumption. If the user can repair that step directly while preserving the rest, the model may avoid repeating the full reasoning process.

This points to an important direction for LLM product design. Better reasoning may not come only from larger models or longer context windows; it may also come from interfaces that make human feedback precise, structured, and reusable. In education, research assistance, and engineering workflows, users often want the model to continue from a corrected draft rather than generate a completely new solution.

The paper summary does not provide every implementation detail or boundary condition, so questions remain about usability, editing effort, and robustness across task types. Still, the core idea addresses a real pain point in current reasoning-model workflows: correction should not always mean starting over. It can mean preserving what is right, fixing what is wrong, and guiding the model forward more efficiently.

Source: arXiv