SIGNAL//DESK
AI/MLsrc: curated AI glossary

Reasoning model

An AI that takes a moment to 'think' and outline its steps before giving you an answer, similar to how a student shows their work on a math problem to avoid making mistakes.

A language model architecture configured to generate hidden or explicit chain-of-thought tokens prior to the final response, effectively utilizing increased inference-time compute to improve accuracy on complex, multi-step tasks.

An LLM optimised to produce intermediate reasoning (and spend additional inference compute) before its final answer, improving performance on logic, planning, and multi-step problems.

evolution

  1. 2017-06 · history
    Transformer Architecture

    The 'Attention Is All You Need' paper established the foundational architecture for modern large language models.

  2. 2022-06 · history
    Chain-of-Thought Prompting

    Researchers introduced techniques to prompt models to generate intermediate reasoning steps, significantly improving performance on complex tasks.

  3. 2023-01 · history
    STaR (Self-Taught Reasoner)

    A framework was introduced allowing models to iteratively improve their reasoning capabilities by generating and filtering their own rationales.

  4. 2024-09 · history
    OpenAI o1 Release

    The first major commercial deployment of a model explicitly trained to perform 'inference-time compute' by processing multi-step reasoning chains before outputting a final answer.


← all terms