SIGNAL//DESK
AI/MLsrc: curated AI glossary

Chain-of-thought

Chain-of-thought is like asking a student to show their work on a math problem instead of just writing down the answer. By forcing the AI to write out each step of its thinking process, it is much less likely to make a mistake and more likely to get the right result.

Chain-of-thought is a prompting technique that encourages a model to generate a sequence of intermediate reasoning steps before producing a final output. This approach decomposes complex problems into smaller, manageable parts, which significantly improves performance on multi-step reasoning, logic, and arithmetic tasks.

A prompting/decoding technique in which the model generates explicit intermediate reasoning steps before the final answer, improving accuracy on multi-step and arithmetic tasks.

evolution

  1. 2022-01 · history
    Chain-of-Thought Prompting

    Wei et al. introduce the technique of prompting models to generate intermediate reasoning steps to improve performance on complex tasks.

  2. 2022-05 · history
    Zero-Shot CoT

    Kojima et al. demonstrate that models can perform reasoning without few-shot examples by simply adding the prompt 'Let's think step by step'.

  3. 2023-05 · history
    Tree of Thoughts

    Yao et al. generalize CoT by allowing models to explore multiple reasoning paths and backtrack, enabling complex problem-solving.

  4. 2024-09 · history
    OpenAI o1 Series

    OpenAI releases models explicitly trained with reinforcement learning to perform extensive internal chain-of-thought processing before outputting a response.


← all terms