A new agent reasoning framework called ReWOO has been developed, which significantly reduces the number of calls to large language models (LLMs) for multi-hop questions. Unlike traditional ReAct frameworks that require a model call for each step of reasoning and re-send the entire transcript, ReWOO separates the process into three roles: Planner, Worker, and Solver. The Planner makes a single LLM call to create a complete plan upfront, the Worker executes the plan using code without further model interaction, and the Solver makes a final LLM call to produce the answer. This approach drastically cuts down LLM calls from k+1 to just 2, regardless of the number of reasoning hops, leading to substantial savings in tokens and reduced latency, though it sacrifices the ability to dynamically adapt to unexpected tool results. AI
IMPACT This framework could significantly reduce operational costs and latency for AI agents handling complex, multi-step queries.
RANK_REASON The item describes a new framework for AI agents that optimizes LLM calls, which is a product/tool development.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
