Chain-of-Thought Prompting: Key Papers and Variants

A reading list and short notes on Chain-of-Thought (CoT) prompting and its variants.

Question Decomposition Improves the Faithfulness of Model-Generated Reasoning

Decomposing a question into smaller sub-questions and answering each separately produces reasoning that is more faithful to the model’s actual computation — i.e., the stated reasoning is more likely to reflect the actual cause of the answer. This addresses a long-standing concern with CoT: that the explanation may be post-hoc rationalization rather than genuine reasoning.

Plan-and-Solve Prompting

Plan-and-Solve improves zero-shot CoT without requiring task-specific exemplars. The prompt explicitly instructs the model to first plan (devise a strategy), then solve (execute the plan), separating high-level reasoning from low-level computation. This two-phase structure reduces the rate of computational mistakes that plague single-pass CoT.

Common Themes

Across these variants, the recurring insight is that CoT’s quality depends heavily on:

• Structural decomposition of the reasoning process
• Separation of planning from execution
• Mechanisms that make the reasoning audit-able and faithful, not just plausible