Source: ReWOO: Decoupling Reasoning from Observations (Xu et al., 2023)
Plan all tool calls up front with evidence variables, execute, then solve: about 5x fewer tokens than ReAct on HotpotQA with slightly higher accuracy.
Contents
- What: research preprint, arXiv:2305.18323
- Authors: Binfeng Xu, Zhiyuan Peng, Bowen Lei, Subhabrata Mukherjee, Yuchen Liu, Dongkuan Xu
- Date: 2023-05-23 (v1)
- Original: arXiv:2305.18323
Summary
ReWOO (“Reasoning WithOut Observation”) argues that ReAct is token-hungry: every step stops generation for a tool call, and the stateless API forces each resumed call to resend the instructions, exemplars and the whole growing trajectory. ReWOO splits the work into three modules. A Planner writes the full plan once, each step naming a tool call whose result is bound to a variable (#E1, #E2, …) that later steps can reference. A Worker executes the calls and fills in the variables. A Solver reads the plan plus evidence and writes the answer.
Key claims
- Prompt tokens in an interleaved loop grow with the number of steps, with a quadratic term from the accumulated history; ReWOO sends the context roughly twice (plan and solve).
- HotpotQA with gpt-3.5-turbo: ReAct 40.8% accuracy, 9,795 tokens, 4.97 steps; ReWOO 42.4%, 1,986 tokens, 4.45 steps (about 5x token efficiency, about 4% relative accuracy gain).
- Consistent gains across six public NLP benchmarks and one curated dataset; more robust when tools fail.
- Because planning no longer depends on observations, plan-only data can fine-tune a small model: the authors move reasoning from GPT-3.5 (175B) into a 7B LLaMA.