Deep Interaction Improves LLM Error Correction and Efficiency

This paper introduces Deep Interaction, an innovative method for human-AI collaboration aimed at efficiently correcting errors in large language models (LLMs), particularly those employing Chain-of-Thought (CoT) reasoning. Current methods often involve regenerating entire responses or laboriously flagging errors, which can lead to recurring mistakes or inefficient correction cycles. Deep Interaction addresses this by enabling users to directly modify specific faulty steps within an LLM's original reasoning process. Once a user corrects an erroneous part, the system distills this refined CoT into an optimized prompt. This new prompt then steers the LLM to follow the corrected reasoning path, preserving the accurate steps while rectifying the mistakes. This direct editing capability offers a more precise and less resource-intensive way to guide LLMs. Experimental results demonstrate the effectiveness of Deep Interaction on STEM reasoning tasks. The method achieved over a 25% improvement in correction success rate and reduced token usage by approximately 40% compared to baseline approaches. This suggests a more efficient and user-friendly paradigm for interacting with and refining complex LLM outputs.