OpenAI：2025年OpenAI o3&o4-mini技术报告（英文版）（32页）.pdf

1、OpenAI o3 and o4-mini System CardOpenAIApril 16,20251IntroductionOpenAI o3 and OpenAI o4-mini combine state-of-the-art reasoning with full tool capabilities web browsing,Python,image and file analysis,image generation,canvas,automations,file search,and memory.These models excel at solving complex ma

2、th,coding,and scientific challenges whiledemonstrating strong visual perception and analysis.The models use tools in their chains ofthought to augment their capabilities;for example,cropping or transforming images,searchingthe web,or using Python to analyze data during their thought process.The Open

3、AI o-series models are trained with large-scale reinforcement learning on chains ofthought.These advanced reasoning capabilities provide new avenues for improving the safetyand robustness of our models.In particular,our models can reason about our safety policies incontext when responding to potenti

4、ally unsafe prompts,through deliberative alignment 11.This is the first launch and system card to be released under Version 2 of our PreparednessFramework.OpenAIs Safety Advisory Group(SAG)reviewed the results of our Preparednessevaluations and determined that OpenAI o3 and o4-mini do not reach the

5、High threshold inany of our three Tracked Categories:Biological and Chemical Capability,Cybersecurity,and AISelf-improvement.We describe these evaluations below,and provide an update on our work tomitigate risks in these areas.2Model Data and TrainingOpenAI reasoning models are trained to reason thr

6、ough reinforcement learning.Models in theo-series family are trained to think before they answer:they can produce a long internal chainof thought before responding to the user.Through training,these models learn to refine theirthinking process,try different strategies,and recognize their mistakes.Re