The core development team of this project mainly consists of third- and fourth-year undergraduate students, as well as first-year PhD students from the GAIR research group at Shanghai Jiao Tong University. The project has been guided by leading research scientists in the field of large language models, including those from NYU and MBZUAI.
The specific introductions of the team members are as follows:
Yiwei Qin, Xuefeng Li, Haoyang Zou, Yixiu Liu, Shijie Xia, Zhen Huang, Yixin Ye, Yuxiang Zheng, Ethan Chern, Weizhe Yuan, Zhengzhong Liu, Yuanzhi Li, Pengfei Liu
- [2025/01/14] 🚨 We have released a report on O1 Inference-time Scaling for Medical Reasoning, demonstrating the immense potential of "deep thinking" in medical diagnosis! Check out our report!
- [2024/11/22] 🚨 We have published a new part of our study on O1-distillation, accompanied by in-depth reflections on the underlying behaviors and phenomena. Check our report!
- [2024/10/16] We have released the journey thought training dataset on 🤗 Hugging Face.
- [2024/10/09] 🚨 We have officially released the first Strategic Report on O1 Replication. We introduce a new training paradigm called ‘journey learning’ and propose the first model that successfully integrates search and learning in mathematical reasoning. The search process incorporates trial-and-error, correction, backtracking, and reflection, making this the first effective approach for complex reasoning tasks. If you do find our resources helpful, please cite our paper.
This research explores a novel approach to enhance medical reasoning capabilities in Large Language Models (LLMs) through inference-time scaling. Building upon our previous work on O1 replication through Journey Learning and Distillation (Part1 & Part2), we investigate how extended reasoning time impacts LLMs' performance across various medical tasks. Through comprehensive experiments on medical benchmarks like MedQA, Medbullets, and JAMA Clinical Challenges, we demonstrate that increased inference time leads to significant performance improvements of 6%-11% with just 500 training samples. Notably, our model's differential diagnosis process mirrors the hypothetico-deductive method used by medical professionals, systematically evaluating evidence to narrow down potential conditions. These findings highlight the powerful combination of inference-time scaling and journey learning in advancing LLMs' ability to tackle complex clinical reasoning tasks.
This paper critically examines current methods for replicating OpenAI’s O1 model, focusing on the often-overlooked use of knowledge distillation. While Part 1 explored foundational techniques for O1 replication, this study demonstrates how simple distillation from O1’s API, combined with supervised fine-tuning, can achieve superior performance on complex mathematical reasoning tasks.
Key findings include:
- Superior Performance with Simplicity: A base model fine-tuned on tens of thousands of O1-distilled chains outperformed O1-preview on the AIME with minimal technical complexity.
- Generalization Across Tasks: Despite being trained only on mathematical data, the distilled models showed strong performance in open-domain QA, reduced hallucination, and improved safety, while becoming less susceptible to sycophancy.
- Transparency and Education: The paper emphasizes the need for transparent research practices and prioritizing first-principles thinking over quick performance gains.
Report (Part1) | Dataset | Walnut Plan | Citation
This paper introduces a pioneering approach to artificial intelligence research, embodied in our O1 Replication Journey. In response to the announcement of OpenAI’s groundbreaking O1 model, we embark on a transparent, real-time exploration to replicate its capabilities while reimagining the process of conducting and communicating AI research.
We propose a new approach: “journey learning”. This paradigm goes beyond the traditional focus on specific tasks and emphasizes continuous progress through learning, reflection, and adaptation. AI systems that follow this method can evolve over time, improving their ability to handle real-world complexities. Unlike shortcut learning, journey learning equips AI with the capacity to adapt, backtrack, and refine its understanding, aiming to create more human-like intelligence. This shift to journey learning opens up new possibilities in AI research, enabling the creation of systems that can not only perform tasks but also reason and grow, making them more capable of engaging with humans across various domains.
- We are gradually releasing resources, and more will be available soon. Stay tuned as we continue to share new updates over time!
If you are interested in our project and would like to join us, feel free to send an email to stefanpengfei@gmail.com.
@article{qin2024o1,
title={O1 Replication Journey: A Strategic Progress Report--Part 1},
author={Qin, Yiwei and Li, Xuefeng and Zou, Haoyang and Liu, Yixiu and Xia, Shijie and Huang, Zhen and Ye, Yixin and Yuan, Weizhe and Liu, Hector and Li, Yuanzhi and others},
journal={arXiv preprint arXiv:2410.18982},
year={2024}
}
@article{huang2024o1,
title={O1 Replication Journey--Part 2: Surpassing O1-preview through Simple Distillation, Big Progress or Bitter Lesson?},
author={Huang, Zhen and Zou, Haoyang and Li, Xuefeng and Liu, Yixiu and Zheng, Yuxiang and Chern, Ethan and Xia, Shijie and Qin, Yiwei and Yuan, Weizhe and Liu, Pengfei},
journal={arXiv preprint arXiv:2411.16489},
year={2024}
}
@article{huang2025o1replicationjourney,
title={O1 Replication Journey -- Part 3: Inference-time Scaling for Medical Reasoning},
author={Zhongzhen Huang and Gui Geng and Shengyi Hua and Zhen Huang and Haoyang Zou and Shaoting Zhang and Pengfei Liu and Xiaofan Zhang},
journal={arXiv preprint arXiv:2501.06458},
year={2025}
}
