ReWOO: Decouple planning from tool observations¶
Explicit Planner/Worker/Solver dependencies and evidence aggregation.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Microsoft Research / North Carolina State University / Texas A&M University. |
| First public release | 2023-05-29 |
| Author code | billxbf/ReWOO. |
| Local adapter or method key | rewoo |
| Local reproduction code | src/auto_research/agent_research/ |
Paper method and evidence¶
Instead of reprompting a ReAct planner after every tool response, write a full plan with variable references. A Worker populates tool evidence and a Solver integrates it once at the end.
flowchart LR
A[Plan once with variable references] --> B[Run dependent tool actions]
B[Run dependent tool actions] --> C[Synthesize final answer]Original paper figure¶
Figure shows the plan, worker and solver separation. From the original paper; copyright remains with the authors.
Core equation¶
\[
P=\{(E_i,T_i,a_i(E_{<i}))\}_{i=1}^{n},\qquad
e_i=T_i(a_i(e_{<i})),\qquad y=S(x,P,e_{1:n}).
\]
Reported results¶
Paper reports about 5× token efficiency, about +4 points on HotpotQA, and planner offload from 175B to 7B. No production A/B.
Local reproduction¶
PlanBench mini, 120 episodes: joint success 1.0000, mean cost 3.5000, 120 full plans and 360 tool/worker actions. Metrics. No real search API, multi-size LLM, or HotpotQA generation.