EP044: How ReAct Synergizes Reasoning and Acting

The paper introduces ReAct, a novel paradigm that combines reasoning and acting within Large Language Models (LLMs) to solve diverse language and decision-making tasks. ReAct prompts LLMs to interleave verbal reasoning traces with task-specific actions. This synergy allows the model to dynamically create, track, and adjust high-level action plans through reasoning, while simultaneously using actions to interface with and gather new information from external environments, such as knowledge bases.By grounding its reasoning in external observations, ReAct effectively overcomes prevalent issues like fact hallucination and error propagation that often limit standard chain-of-thought (CoT) reasoning. Empirical evaluations demonstrate that ReAct outperforms state-of-the-art baselines across multiple benchmarks, including question answering (HotpotQA), fact verification (Fever), text-based games (ALFWorld), and web navigation (WebShop). Furthermore, the interleaved thought-action process significantly improves human interpretability and trustworthiness, as users can easily inspect the model's decision-making basis and even correct its behavior on the fly by editing its reasoning traces.