OpenAI o1 introduces reasoning models

o1 "thinks" before answering using RL-trained chain of thought, jumping from 13% to 83% on an IMO qualifying exam.

What happened and why it matters

o1 opened a second scaling axis — test-time compute. Performance improves with more thinking time, and RL on reasoning traces produced large gains on maths, science (GPQA above PhD-expert baseline) and competitive programming.

Sources

• Learning to reason with LLMs ↗ OpenAI · 12 Sep 2024