Part 4/8:
Instead of creating numerous intermediate steps filled with excess information, Chain of Draft allows LLMs to produce "dense" information outputs at each stage. Simply put, models are instructed to summarize their thinking in a concise manner, effectively advancing towards the solution without the burden of extensive reasoning.
The innovation doesn't require any complex modifications or updates to existing models; it only necessitates adjustments in the prompting strategy itself. By simply rephrasing prompts to focus on highlighting essential thinking steps, researchers demonstrated a significant improvement in efficiency.