- https://arxiv.org/pdf/2210.03629.pdf
- Source: arxiv
TL;DR
ReAct interleaves verbal reasoning traces and action steps in a single LLM forward pass, enabling grounded, interpretable agent behavior that outperforms chain-of-thought and action-only baselines.
Summary
Separating reasoning from acting in LLM agents leads to ungrounded thoughts or blind actions. ReAct solves this by having the model produce a reasoning trace before each action and then incorporating the action’s output into the next reasoning step. This dynamic loop outperforms chain-of-thought and action-only baselines on HotpotQA, FEVER, ALFWorld, and WebShop. The interpretable trace structure also enables practical human-in-the-loop correction mid-task, without any fine-tuning.
Key contributions
- Proposes the ReAct prompting paradigm that interleaves reasoning traces and actions in a unified LLM output.
- Demonstrates that grounding reasoning in observed action results reduces hallucination compared to chain-of-thought alone.
- Shows that few-shot ReAct prompts are sufficient for strong performance across both knowledge retrieval and embodied decision tasks.
- Introduces interpretable agent traces as an audit log enabling human inspection and mid-task correction.
When to cite
- When building an LLM agent that needs to alternate between reasoning and tool/environment interaction.
- When justifying why interleaved reasoning and acting outperforms either in isolation.
- When arguing for interpretable agent architectures that support human-in-the-loop oversight.
- When choosing a baseline prompting strategy for knowledge-intensive QA or interactive decision tasks.