Seonglae ChoLLM reasoning (chain-of-thought prompting) and acting (action plan generation) have largely been studied as separate topics. CoT is a static black box that relies only on the model’s internal representations, so it is not grounded in the external world and is therefore vulnerable to factual hallucination and error propagation. Conversely, acting-focused approaches (e.g., WebGPT, SayCan) do not leverage higher-level reasoning or working memory for abstract goals. This paper shows that interleaving the generation of these two abilities produces synergistic benefits.
The core idea is to expand an agent’s action space into the language space.
