ReAct: Synergizing Reasoning and Acting in Language Models

Type: paper Tier: 2 (Foundational Paper) Author(s): Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao Date: 2022 (submitted 2022-10-06) URL: https://arxiv.org/abs/2210.03629 Accessed: 2026-06-01

Why This Source Matters

ReAct is the foundational paper for the tool-use loop with reasoning — the structure underneath every modern coding copilot and agent. Before ReAct, prompting split into two camps: reasoning-only (chain-of-thought, which thinks but can't look anything up, so it hallucinates) and acting-only (call tools, but with no plan). ReAct's contribution is to interleave the two: the model alternates between a thought (reasoning trace) and an action (a tool call), observing each action's result before the next thought. This is exactly the loop Project 06 builds — think about what file to read, read it, observe, think again, answer — and it is the direct conceptual ancestor of Anthropic-style tool use.

Key Claims

The central idea: interleave reasoning and acting

• "We explore the use of LLMs to generate both reasoning traces and task-specific actions in an interleaved manner." The model produces a Thought → Action → Observation trajectory, repeated until it can answer.

The two directions reinforce each other

• "Reasoning traces help the model induce, track, and update action plans as well as handle exceptions, while actions allow it to interface with external sources, such as knowledge bases or environments, to gather additional information." Reasoning makes acting purposeful; acting grounds reasoning in real observations.

The action space includes external tools

• Actions interface with external systems — in the paper, "a simple Wikipedia API" and interactive environments — letting the model "retrieve real information rather than relying solely on parametric knowledge." (In Project 06, the action space is read_file, list_directory, search_code.)

It beats reasoning-only and acting-only

• ReAct surpasses both baselines. On decision-making benchmarks ALFWorld and WebShop it achieves "34% and 10% absolute success rate" improvements over imitation/RL methods, "with one or two in-context examples."

It reduces hallucination and is more interpretable

• ReAct "overcomes issues of hallucination and error propagation prevalent in chain-of-thought reasoning" (because actions fetch ground truth), and it generates "human-like task-solving trajectories that are more interpretable than baselines" (you can read the thoughts).

Relevant To

• concepts: [react, agentic-loop, tool-use, reasoning-trace, hallucination-reduction, observation-action]
• projects: [06-ai-coding-copilot, 08-ai-agent]

Notes

• ReAct = the loop's grammar. Thought (reason about what's needed) → Action (call a tool) → Observation (tool result) → repeat → Answer. Project 06's run_agent is a minimal ReAct loop; the model's "thought" is whatever text it emits alongside (or before) a tool call.
• Why acting reduces hallucination: a chain-of-thought-only model asked "what does config.py set the default model to?" will guess; a ReAct model reads the file first. This is the single best argument for tool use over a clever prompt — and the failure mode the learner should reproduce in M6.
• Why reasoning makes acting terminate: acting-only loops thrash (call the same tool, never decide they're done). The reasoning trace is what lets the model conclude "I now have enough to answer" — connects to why the loop needs both a reasoning step and a max_steps cap (sources/official-docs/anthropic-tool-use.md).
• Modern tool-use APIs (Anthropic tool_use/tool_result, OpenAI tool_calls) are ReAct operationalized as a protocol: the "Action" is the structured tool call, the "Observation" is the tool_result you feed back.