Chain-of-Thought (CoT) prompting asks a model to generate intermediate reasoning steps before producing a final answer.
Why It Matters for Agents
Agents often need to decide which tool to call, what to do next, and when to stop. CoT provides a mechanism for deliberation that can make multi-step decisions more reliable.
Common Variants
- Zero-shot CoT: add a trigger like “Let’s think step by step.”
- Few-shot CoT: include examples that show full reasoning traces.
- Auto-CoT: automatically generate diverse reasoning traces, then reuse them as demonstrations.
Trade-offs
- Can improve reliability on multi-step tasks.
- Can increase latency and cost (more tokens).
- Reasoning traces may be unfaithful (they are not guaranteed to reflect true internal computation).
Further Reading
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models --- Wei et al. (2022)