ReAct: Synergizing Reasoning and Acting in Language Models
Abstract
While large language models (LLMs) have demonstrated impressive capabilities across tasks in language understanding and interactive decision making, their abilities for reasoning (e.g. chain-of-thought prompting) and acting (e.g. action plan generation) have primarily been studied as separate topics. In this paper, we explore the use of LLMs to generate both reasoning traces and task-specific actions in an interleaved manner, allowing for greater synergy between the two: 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. We apply our approach, named ReAct, to a diverse set of language and decision making tasks and demonstrate its effectiveness over state-of-the-art baselines, as well as improved human interpretability and trustworthiness over methods without reasoning or acting components. Concretely, on question answering (HotpotQA) and fact verification (Fever), ReAct overcomes issues of hallucination and error propagation prevalent in chain-of-thought reasoning by interacting with a simple Wikipedia API, and generates human-like task-solving trajectories that are more interpretable than baselines without reasoning traces. On two interactive decision making benchmarks (ALFWorld and WebShop), ReAct outperforms imitation and reinforcement learning methods by an absolute success rate of 34% and 10% respectively, while being prompted with only one or two in-context examples. Project site with code: https://react-lm.github.io
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
ReAct introduces a prompting paradigm that interleaves verbal reasoning traces with task-specific actions, enabling LLMs to dynamically plan, interact with external environments, and self-correct, thereby establishing a foundational framework for modern LLM agents. The paper's primary strength lies in its conceptual clarity and empirical robustness across diverse benchmarks, demonstrating that simple structural modifications to LLM prompting can yield substantial gains in reliability, interpretability, and task success. While it does not introduce novel architectures or training objectives, its influence on the field is profound, catalyzing a wave of research into tool-augmented reasoning, agent frameworks, and interactive LLM systems. The rigorous evaluation, transparent methodology, and clear articulation of failure modes make it a highly impactful contribution that will remain a core reference for agent-based AI research.
Methodology
The paper proposes a remarkably elegant prompting framework that interleaves Chain-of-Thought (CoT) style reasoning traces with explicit action calls and environmental observations. By structuring the LLM's generation into a "Thought → Action → Observation" loop, the method creates a closed feedback cycle where reasoning guides tool selection, and external feedback grounds subsequent reasoning steps. The approach requires no parameter updates, relying entirely on in-context learning with 1-2 demonstrations. While conceptually straightforward, the explicit design choice to synergize reasoning and acting addresses two critical failure modes of prior methods: CoT's tendency to hallucinate without grounding, and Act-only approaches' inability to recover from errors or plan multi-step strategies. The methodology is highly modular and model-agnostic, making it trivially adaptable to any sufficiently capable LLM.
Results & evaluation
The experimental design is rigorous and spans four distinct, challenging benchmarks: multi-hop QA (HotpotQA), fact verification (FEVER), and two interactive decision-making environments (ALFWorld, WebShop). ReAct consistently outperforms both CoT-only and Act-only baselines, with particularly dramatic gains in interactive settings (e.g., +34% absolute success rate on ALFWorld). The ablation studies effectively isolate the contribution of reasoning traces, demonstrating their critical role in exception handling and plan updating. The use of both PaLM and GPT-3 strengthens the claim of generalizability, and the qualitative analysis of generated trajectories provides compelling evidence of improved interpretability and human-like problem-solving behavior.
Reproducibility
High for a prompting-based study. The authors provide all prompt templates in the appendix and release the GPT-3 evaluation code. While exact reproduction on PaLM is limited by its closed-source nature, the methodology's reliance on standard prompting paradigms ensures that results can be readily replicated on any modern open or closed LLM. The paper transparently acknowledges this limitation and provides sufficient implementation details to mitigate it.
Limitations
As a zero/few-shot prompting technique, ReAct's performance ceiling is strictly bounded by the base model's capabilities, particularly in long-context retention and complex logical deduction. The interleaved format significantly increases token consumption and inference latency compared to direct action generation. Additionally, while error recovery is improved, the method lacks explicit mechanisms for backtracking or global state tracking, which can lead to compounding errors in highly complex, long-horizon tasks. The approach also assumes well-structured, deterministic external APIs, limiting immediate applicability to noisy or partially observable real-world environments.
Broader impact
ReAct has fundamentally shaped the trajectory of LLM agent research, establishing the reasoning-acting loop as a de facto standard for tool-augmented language models. Its paradigm directly enables applications in automated research assistants, web navigation, code execution, and embodied AI. The paper responsibly addresses safety concerns, noting the risks of connecting LLMs to open environments and advocating for constrained action spaces during research. By making LLM behavior more transparent and diagnosable, ReAct also contributes to the broader push for interpretable and trustworthy AI systems. ReAct introduces a prompting paradigm that interleaves verbal reasoning traces with task-specific actions, enabling LLMs to dynamically plan, interact with external environments, and self-correct, thereby establishing a foundational framework for modern LLM agents. The paper's primary strength lies in its conceptual clarity and empirical robustness across diverse benchmarks, demonstrating that simple structural modifications to LLM prompting can yield substantial gains in reliability, interpretability, and task success. While it does not introduce novel architectures or training objectives, its influence on the field is profound, catalyzing a wave of research into tool-augmented reasoning, agent frameworks, and interactive LLM systems. The rigorous evaluation, transparent methodology, and clear articulation of failure modes make it a highly impactful contribution that will remain a core reference for agent-based AI research.
Curator note
Yao et al.; interleaved reasoning and tool use; foundation of agents
Author of this paper? Send a correction.