Summary: Yao et al. (2023) — ReAct¶
Full citation: Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.
Key Contribution¶
This paper introduces ReAct, the prompting paradigm that became the standard implementation pattern for LLM agents. ReAct interleaves reasoning traces (thoughts expressed in natural language) with actions (tool calls to external environments) and observations (results returned from those actions). The key insight: reasoning and acting are synergistic — reasoning helps the agent plan and interpret observations, while actions ground reasoning in external evidence, reducing hallucination.
The ReAct Pattern¶
At each step, the agent generates one of:
- Thought: free-form reasoning — decomposing the goal, extracting information from observations, performing commonsense inference, tracking progress, reformulating search queries, handling exceptions
- Action: a structured call to an external tool (e.g., search[entity], lookup[string], finish[answer])
- Observation: the result returned by the environment (not generated by the model)
The thought-action-observation cycle repeats until the agent reaches a conclusion or a stopping condition. This is the pattern described conceptually in perceive-reason-act-loop and used by virtually all modern agent frameworks.
Key Empirical Findings¶
ReAct vs Chain-of-Thought (CoT) — failure mode comparison (Table 2):
| Failure Mode | ReAct | CoT |
|---|---|---|
| Hallucination (fabricated facts/traces) | 14% | 56% |
| Reasoning error (repetitive loops, failed recovery) | 47% | 16% |
| Correct reasoning and facts | 94% (success) | 86% (success) |
The critical finding for safety professionals: CoT's dominant failure mode is hallucination (56%), because reasoning from internal knowledge alone is not grounded in external evidence. ReAct dramatically reduces hallucination to 14% by grounding reasoning in retrieved observations. However, ReAct introduces its own failure mode — reasoning errors where the agent gets stuck in repetitive action patterns or fails to recover from non-informative search results.
Best overall approach — ReAct + CoT-SC: Combining ReAct with chain-of-thought self-consistency (multiple reasoning samples with majority vote) outperforms either alone. The model uses internal reasoning when confident and falls back to retrieval when not, supporting the hybrid-decision-pipeline concept.
Decision-making tasks: On ALFWorld (household navigation) and WebShop (web shopping), ReAct outperforms action-only baselines by 33-90% relative improvement. Without reasoning traces, agents fail to decompose goals, track subgoals, or handle exceptions.
Interpretability: ReAct traces are human-readable — a reviewer can inspect the thought that led to each action and verify whether the reasoning was sound. This connects to opacity-and-explainability: while the model's internal computation remains opaque, the ReAct trace provides a structured rationale that supports auditability.
ReAct-Specific Failure Modes¶
Three failure modes are characteristic of the ReAct pattern: 1. Repetitive action loops: the agent calls the same search repeatedly without reformulating 2. Search result errors: non-informative retrieval derails the reasoning chain (23% of failures) 3. Reasoning errors: the model fails to recover from errors in its own reasoning trace (47% of failures)
These failure modes are relevant to tool-calling — they show how tool-calling failures propagate through the agent loop, and why circuit breakers and fallback strategies are needed.
Relevance to This Wiki¶
ReAct is the architectural pattern underlying virtually every LLM agent system discussed in this wiki. The empirical comparison with CoT provides the strongest evidence that external grounding reduces hallucination — the central argument for retrieval-augmented-generation and knowledge-graphs. The failure mode analysis shows that even with grounding, agent reasoning can get stuck in unproductive patterns, motivating the governance and monitoring mechanisms described in governance-gates and multi-agent-coordination-failures.