Chain-of-Thought in Language Models

• Chain-of-Thought is a prompting strategy that explicitly generates intermediate reasoning steps to enhance model predictions.
• It improves compositional robustness and systematic generalization by making latent reasoning processes transparent.
• The approach supports verification of model behavior and has spurred innovations like symbolic scaffolding and conceptual reasoning.

Chain-of-Thought (CoT) denotes a class of prompting, learning, and interpretability techniques for large language and vision-LLMs that structures reasoning as an explicit sequence of intermediate steps. By forcing models to articulate partial deductions before a final prediction, CoT can unlock latent reasoning capacities, enhance systematic generalization, improve compositional robustness, and provide a tractable interface for monitoring and verification of model behavior. The CoT paradigm has also spurred rich theoretical analysis, architectural innovations, and the proposal of extensions such as symbolic scaffolding, conceptual chain-of-thought, and neural search over reasoning paths.

1. Formal Definitions and Theoretical Underpinnings

At its core, Chain-of-Thought prompting transforms the prediction task from a direct mapping $P(y|x)$ (input-to-answer) into a two-stage generation:

• The model is prompted to sample an explicit intermediate reasoning trace $z = (z_