Advancing AI Post-Training Research

Rafael Mitkov Rafailov, the “DPO guy,”  is a researcher and practitioner in machine learning best known for inventing Direct Preference Optimization (DPO). Since its release, DPO has redefined how large language models are aligned with human and synthetic feedback, replacing heavyweight RLHF pipelines with a simple, scalable alternative. DPO is now at the core of instruction-tuned models from leading labs like Meta’s LLaMA, Mistral’s open-weight releases. Beyond DPO, Rafael’s research covers reinforcement learning, imitation learning, and preference modeling, where he develops alignment methods that are both theoretically elegant and deployable at scale.

Stella Biderman, Founding Advisor of SynthLabs, leads EleutherAI, where she drives open-source AI research and development. She has been instrumental in creating widely accessible language models and datasets, advancing both the democratization of AI research and progress in the field.