Deep Interaction: An Efficient Human-AI Interaction Method for Large Reasoning Models
Merged summary
TL;DR — "Deep Interaction" is a human-in-the-loop mechanism that lets users directly edit a large reasoning model's Chain-of-Thought to fix faulty steps, then distills the corrected reasoning into a prompt that steers the model, improving error correction while cutting token usage.
- Targets a common failure mode: re-generation or vague follow-up flagging often reproduces the same reasoning mistakes.
- Users edit the original response to correct erroneous steps while preserving valid reasoning, rather than regenerating from scratch.
- The edited CoT is refined into a "distilled prompt" that guides the LLM along the corrected path.
- Reported gains: >25% higher correction success rate and ~40% lower token usage on STEM reasoning tasks vs. baselines (authors' claims).
Sources (1)
Deep Interaction: An Efficient Human-AI Interaction Method for Large Reasoning Models
TL;DR — "Deep Interaction" is a human-in-the-loop mechanism that lets users directly edit a large reasoning model's Chain-of-Thought to fix faulty steps, then distills the corrected reasoning into a prompt that steers the model, improving error correction while cutting token usage.
- Targets a common failure mode: re-generation or vague follow-up flagging often reproduces the same reasoning mistakes.
- Users edit the original response to correct erroneous steps while preserving valid reasoning, rather than regenerating from scratch.
- The edited CoT is refined into a "distilled prompt" that guides the LLM along the corrected path.
- Reported gains: >25% higher correction success rate and ~40% lower token usage on STEM reasoning tasks vs. baselines (authors' claims).