OpenAI o1 System Card

TL;DR

The OpenAI o1 system employs large-scale reinforcement learning with chain of thought reasoning to enhance safety, robustness, and risk management, achieving state-of-the-art results on safety benchmarks.

Contribution

This work introduces the o1 model series with advanced reasoning capabilities and safety evaluation methods, highlighting the importance of alignment and risk mitigation.

Findings

State-of-the-art safety performance on risk benchmarks

Models can reason about safety policies in context

Emphasizes need for robust alignment and stress-testing

Abstract

The o1 model series is trained with large-scale reinforcement learning to reason using chain of thought. These advanced reasoning capabilities provide new avenues for improving the safety and robustness of our models. In particular, our models can reason about our safety policies in context when responding to potentially unsafe prompts, through deliberative alignment. This leads to state-of-the-art performance on certain benchmarks for risks such as generating illicit advice, choosing stereotyped responses, and succumbing to known jailbreaks. Training models to incorporate a chain of thought before answering has the potential to unlock substantial benefits, while also increasing potential risks that stem from heightened intelligence. Our results underscore the need for building robust alignment methods, extensively stress-testing their efficacy, and maintaining meticulous risk management protocols. This report outlines the safety work carried out for the OpenAI o1 and OpenAI o1-mini models, including safety evaluations, external red teaming, and Preparedness Framework evaluations.