Deep Interaction: An Efficient Human-AI Interaction Method for Large Reasoning Models
arXiv paper 'Deep Interaction: An Efficient Human-AI Interaction Method for Large Reasoning Models' proposes a method called Deep Interaction that allows users to directly edit the raw response of large language models to correct reasoning errors, and distills the edited chain-of-thought into prompts to guide the model. Experiments show that on STEM task reasoning, this method improves correction success rate by over 25% compared to baseline methods and reduces token usage by about 40%.
This method achieves precise error correction by directly editing the raw response and distilling the corrected chain-of-thought into prompts, while preserving correct reasoning steps.
This research addresses the inefficiency of correcting reasoning errors in large language models by proposing a human-AI interaction method that may improve model usability in complex tasks.
By reducing token usage (about 40%) and improving correction success rate (over 25%), this method may lower inference costs and increase user satisfaction.
Future work could focus on the generalization ability of this method across more tasks and models, and whether it will be integrated into practical products.