The emergence of Chain-of-Thought (CoT) reasoning has significantly enhanced large language models (LLMs) in tackling complex, multi-step tasks. However, when errors occur, current interaction methods typically involve re-generating responses which may lead to repeated mistakes, or users must painstakingly flag faulty steps that might result in similar recurring errors.
To address this, we propose an efficient human intervention mechanism called Deep Interaction for precisely correcting reasoning errors in LLMs.
Our approach allows for direct editing of the original response, correcting erroneous parts while preserving accurate reasoning steps. We refine the edited CoT into a distilled prompt, guiding the LLM along the corrected reasoning path. Experimental results demonstrate that our method achieves over a 25% improvement in correction success rate and reduces token usage by approximately 40% on STEM reasoning tasks compared to baseline approaches.
Blogger's Review: The Deep Interaction method shows significant advantages in correcting reasoning errors in LLMs, especially when handling complex tasks. The direct editing approach greatly reduces user workload, and its improvements in success rates and token usage indicate a promising future for efficient human-AI interaction.