简体中文 · English

ReAct

The local implementation alternates Thought → Action → Observation and records actual reasoning steps and actions. It is not a direct answer wrapped as ReAct.

Paper information

Field	Details
Paper	ReAct: Synergizing Reasoning and Acting in Language Models
Institution	Princeton University / Google Research
First public release	2022-10-06
Author code	Released
Local adapter or method key	react
Local reproduction code	src/auto_research/agent_research/

Original paper summary

Background and main changes

Chain-of-thought alone can hallucinate from closed-book knowledge; action-only agents lack planning and state tracking. ReAct alternates natural-language reasoning with environment actions and adds each observation to the next context. Reasoning can correct errors and actions can obtain external facts.

flowchart LR
    Q["Task and current observation"] --> T["Thought: update plan"]
    T --> A["Action: use tool or environment"]
    A --> O["Observation: fact or state"]
    O --> T
    T --> F["Final answer"]

Key figure from the paper

Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations

\[ \tau=(o_0,t_1,a_1,o_1,\ldots,t_T,a_T,o_T),\qquad (t_i,a_i)\sim\pi_\theta(\cdot\mid o_{\le i-1},t_{<i},a_{<i}). \]

The contribution is not a new loss; it puts the reasoning trace and executable actions in one trajectory.

Offline and online results

The paper improves success over imitation/RL baselines by 34 and 10 percentage points on ALFWorld and WebShop, respectively. It uses the Wikipedia API to reduce hallucination on HotpotQA and FEVER. No production online A/B test is reported.

Local reproduction

ScaleMCP mini uses 120 fixed episodes and seed 42. The implementation performs one reasoning step before each required tool, then records the action and observation.

Metric	ReAct
Joint success	1.0000
Average cost	3.0000
Reasoning steps / actions	360 / 360

auto-research agent-eval --method react --benchmark scalemcp-mini \
  --episodes 120 --memory-size 24 --seed 42

Stable metrics: classic-agent-mini-suites-seed42.json.

Scope

The alternating trace and tool-action counts are retained, but observations come from a deterministic mini-suite. It does not call Wikipedia, WebShop, or a paid LLM. The local 1.0 only verifies the execution loop and is not comparable with the paper's results.

Unified L2.1 isolated capability result

On the isolated toolroute-l2.1-v1 test split (60 episodes/seed, seeds 42/43/44), ReAct reaches joint success 0.6833, plan-step F1 0.7683, and failure recovery 0.6667. The policy interface has no reference answer, plan, or route, and no guide. See metrics and the L2.1 isolated capability protocol.