Deep Interaction: Fixing AI Reasoning Errors by Editing, Not Regenerating

Hefeng Zhou, Jinxuan Zhang, Jiong Lou, Yuxin Liu, Chaochao Lu, Jingjing Qu, Jie Li · July 15, 2026

Large language models (LLMs) that use chain-of-thought reasoning can solve complex problems step by step, but they often make mistakes. Current fixes require users to flag errors and have the model regenerate a new answer, which may repeat the same mistake or waste tokens. Researchers propose Deep Interaction, a method that lets users directly edit the erroneous part of the model's response while keeping correct steps intact. The edited chain-of-thought is then distilled into a concise prompt that guides the model along the corrected path. In STEM tasks, Deep Interaction improves correction success by over 25% and reduces token usage by about 40% compared to baseline methods. This approach could make human-AI collaboration more efficient, especially in fields like math, science, and programming where precise reasoning is critical. By saving time and computational resources, it also lowers costs for users and developers.