Chain-of-Thought

Chain-of-Thought (CoT) is a prompting and reasoning technique for large language models, first introduced around 2022 by researcher Jason Wei, then at Google and now at OpenAI . The core idea is that models perform better on complex tasks when they generate detailed intermediate reasoning steps rather than jumping directly to an answer—an approach later simplified by the "let's think step by step" prompting method .

In its standard form, CoT produces explicit, readable chains of language tokens that trace the model's reasoning process, making it useful for tasks in mathematics, coding, and logical inference . The technique has since branched into two main directions: **explicit** variants (including Tree-of-Thought and Graph-of-Thought) that structure reasoning with verifiable, decomposable steps ; and **implicit** approaches that embed reasoning in continuous latent spaces rather than natural language, such as Meta AI's Coconut method . Per a 5Y Capital analysis from 2023, CoT likely works by activating more complex "task circuits" in a model's weights compared to direct answering .

OpenAI's o1 model brought renewed attention to CoT by combining it with reinforcement learning: rather than training on human-written reasoning steps, o1 appears to learn through reward signals applied to its own generated reasoning paths, with step-level granularity . As a ZhenFund-transcribed DeepMind researcher noted, the raw chain often contains human-like hesitation markers ("hmm," "let me see") that suggest the model is doing genuine reflective processing . However, as Yunqi Capital cautioned in September 2024, CoT does not eliminate hallucinations, and OpenAI has chosen not to expose the raw chain to users for usability reasons .