pith:VTZ5TICI

Prompting a frozen LLM (PaLM-540B, also GPT-3) to emit interleaved free-form reasoning traces and domain actions in one trajectory outperforms reasoning-only (CoT/CoT-SC) and acting-only baselines, and on ALFWorld and WebShop with only 1–2 in-context demonstrations beats imitation and imitation+RL agents trained on 10³–10⁵ task instances by 34 and 10 absolute success-rate points, respectively.

That the gains are attributable to the ReAct prompting pattern rather than to prompt-engineering and example-selection effects on a small set of held-out tasks. Prompts are hand-authored per task type, ALFWorld results are reported as best-of-6 prompt permutations against a beam-search BUTLER baseline, and HotpotQA EM (27.4) is below CoT (29.4) and far below supervised SoTA, so the headline "outperforms" claim depends on which benchmark and which baseline. The decision-task comparison also conflates a 540B LLM with much smaller imitation/RL agents, leaving the contribution of scale vs. method partly unidentified.

Interleaving generated reasoning traces with tool/environment actions in a single LLM prompt yields large gains over chain-of-thought-only and action-only baselines on QA, fact-checking, and text-game/web-shopping benchmarks.