ReAct: Synergizing Reasoning and Acting in Language Models

An LLM's abilities for REASONING (chain-of-thought) and for ACTING (action-plan generation) had been studied as separate topics, so reasoning traces could not ground themselves in external information and actions could not be error-checked.

Interleaves generation of REASONING TRACES and TASK-SPECIFIC ACTIONS in a single alternating sequence. Reasoning traces let the model induce, track, and update action plans and handle exceptions; actions let it INTERFACE WITH EXTERNAL SOURCES such as knowledge bases or environments to gather information that then conditions the next trace. The alternation is the mechanism - each step's observations feed the next reasoning step, which is what closes the loop.

No inference-speed claim. The reported results are quality: on HotpotQA and FEVER, ReAct overcomes the hallucination and error-propagation problems prevalent in chain-of-thought reasoning by interacting with a simple Wikipedia API; on ALFWorld and WebShop it 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.

ADOPTION: in-production - the reasoning/action interleaving pattern is the direct ancestor of essentially every deployed tool-using agent, and needs no engine support, only a tool-calling loop. VENUE: ICLR 2023 - the arXiv comment states 'v3 is the ICLR camera ready version', so venue rests on the primary source. WHY IT IS IN THE INFERENCE-COST SLICE despite making no efficiency claim: ReAct MULTIPLIES generation calls - each thought/action/observation cycle is another decode pass, and the observation tokens are prompt tokens that re-enter prefill on the next turn. Combined with the fact that ReAct's traces share a growing prefix across turns, this makes ReAct the canonical case where KV-cache prefix reuse (pap-eff-sglang-structured-program-runtime's RadixAttention) pays off, and where the repo's agentic-KV-reuse interest lives. That prefix-reuse arithmetic is not in the paper and is recorded here as the inference-cost reading, not as a paper claim. AFFILIATIONS FILLED 2026-10-04 (paper-verification pass): from the Semantic Scholar graph API (paper/arXiv:2210.03629), which is the only machine-readable source that carries author affiliations — the arXiv Atom feed returns author names only, and no ORCID is exposed by either API. One known upstream typo corrected: "Tel Aviv Univesity" -> "Tel Aviv University". S2 affiliation coverage is partial (120 of 249 corpus records carry any), so these lists may be incomplete rather than exhaustive. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: iclr has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.