Researched
Chain-of-Thought Prompting
Jason Wei and colleagues showed in January 2022 that a language model solves maths problems far better when it writes out its steps first; this habit became the core of today's reasoning models.
Open in the interactive tree →In their January 2022 paper, a 540-billion-parameter model shown eight worked examples with step-by-step reasoning reached the best score then known on the GSM8K school-maths test, beating a fine-tuned GPT-3 with a checker. In May 2022 Kojima and colleagues found that simply adding "Let's think step by step" raised one arithmetic score from 17.7% to 78.7%. OpenAI's o1 (September 2024) and DeepSeek-R1 (January 2025) later used reinforcement learning to make models produce long chains of thought on their own, which is the idea behind reasoning models.
Prerequisites
- Pretrained Language Models2018Few-shot prompting needs a model pretrained on huge amounts of text.
- Neural Scaling Laws2020The method worked on very large models, such as one with 540 billion parameters.
Unlocks
- Reasoning Models2024Reasoning models train the step-by-step habit with reinforcement learning.