CoT for Action Grounding

Use natural-language chain-of-thought to map instructions onto discrete action sequences when the hard part is choosing and ordering grounded steps.

Commonsense chain-of-thought prompting is not only for math: linguistic intermediate steps help tasks that need physical or social background knowledge. On SayCan-style benchmarks, CoT helps map a natural-language instruction to a discrete action sequence — the same shape as tool/plan selection in agents. Wei et al. report large PaLM 540B CoT gains on StrategyQA, Sports Understanding, Date Understanding, and SayCan, while CSQA gains stay minimal (prior-knowledge MC that may not need multi-step decomposition).

Harness takeaway: prefer CoT (or forced reasoning before tools / plan-and-solve prompting) when the failure mode is multi-hop strategy or action ordering — not when the bottleneck is a single fact lookup. Still validate plans with plan-validate-execute and world-model-augmented planning; a CoT action list is not an executable world model. Gate with when chain of thought helps and offline prompt evaluation.

---

Sources

• Chain-of-Thought Prompting (Wei et al.), §4