AI BriefingLearning LLM Reasoning Ability
Key point
OpenAI has unveiled a new model, o1-preview, which maximizes reasoning ability through reinforcement learning.
Details
OpenAI has unveiled a new model, o1-preview. This model demonstrates human expert-level performance in fields requiring advanced reasoning, such as coding (Codeforces), mathematics (AIME), and science (GPQA).
o1 learns to think productively by leveraging Chain of Thought through large-scale Reinforcement Learning algorithms. In particular, it shows the characteristic of continuously improving performance as both the amount of compute invested during training (train-time compute) and the time spent during inference (test-time compute) increase.
Key benchmark results are as follows:
- AIME (Mathematics): While GPT-4o showed an average accuracy of 12%, o1 recorded 74% accuracy with a single sample and 93% with re-ranking, reaching a level equivalent to the top 500 in the US Mathematical Olympiad.
- GPQA (Science): It became the first model to surpass PhD-level human performance on the GPQA-diamond benchmark, which is at an expert level in chemistry, physics, and biology.
- MMMU (Visual Perception): With visual perception capabilities enabled, it recorded 78.2%, achieving a level competitive with human experts.
When solving problems, o1 undergoes a human-like thought process, recognizing and correcting its own errors, simplifying complex steps, or trying new approaches.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.