Skip to content

ReAct

ReAct (Reasoning + Acting) is covered in the RAG and Agents repo under Agent Fundamentals — it is the foundational agentic pattern and is documented there in full.

Quick summary:

ReAct interleaves reasoning traces (Thought) with actions (Action) and their results (Observation) in a loop:

Thought: I need to find the population of Tokyo.
Action: search("Tokyo population 2024")
Observation: Tokyo population is approximately 13.96 million (city) / 37.4 million (metro).
Thought: I have the answer.
Action: finish("13.96 million in the city proper")

Key properties:

  • Thought steps are not executed — they are internal scratchpad reasoning
  • Actions call external tools (search, calculator, code executor, database)
  • The loop continues until the agent emits a finish action
  • Introduced by Yao et al. (2023); shown to outperform CoT-only and Act-only baselines on HotpotQA, Fever, AlfWorld

In the context of alignment: ReAct reasoning traces serve as interpretable audit logs — each decision step is visible, which supports human oversight and reward model training on reasoning quality.

Reference: Yao et al. (2023) — "ReAct: Synergizing Reasoning and Acting in Language Models." ICLR 2023.