The authors propose Deep Interaction, an efficient human-AI interaction method designed to correct reasoning errors in large language models without requiring full response regeneration. This mechanism allows users to directly edit the original Chain-of-Thought response, preserving accurate steps while fixing mistakes.
- The approach refines the edited CoT into a distilled prompt to steer the model along the corrected path.
- It achieves over a 25% improvement in correction success rate on STEM tasks compared to baselines.
- Token usage is reduced by approximately 40% relative to baseline approaches.
This method addresses the inefficiency of current interaction patterns where users must laboriously flag faulty steps or accept recurring errors.