Unknown architecture and mechanisms of OpenAI’s o1 Large Reasoning Model

Characterize the architecture, operational mechanisms, and internal capabilities of OpenAI’s o1 (Strawberry) Large Reasoning Model that differentiate it from standard autoregressive Large Language Models, including the structures employed during pretraining and inference.

Background

The paper distinguishes Large Reasoning Models (LRMs) like OpenAI’s o1 from prior LLMs, noting that o1 appears fundamentally different in nature. However, technical details about its architecture and operational mechanisms are not disclosed by OpenAI, hindering rigorous evaluation, interpretability, and the development of appropriate benchmarking tools.

This unknown limits researchers’ ability to assess how o1 achieves its reasoning capabilities, including whether and how reinforcement learning, rollout-style inference, or private Chain-of-Thought traces are integrated, and complicates comparisons with classical planners and LLM-Modulo systems.