Open and Reliable Language Model Adaptation
Presented by Faeze
Faeze‘s research focuses on understanding and improving the capabilities of language models in real-world, dynamic settings. She develops resource-efficient algorithms for constrained reasoning and generation, investigates how training data and alignment strategies shape model behavior, and designs robust evaluation frameworks to capture emergent capabilities. Her broader goal is to build human-centered AI systems that are reliable, adaptive, and safe across diverse applications.
Abstract
In this talk, Faeze explores two crucial frontiers in AI development: democratizing language model adaptation and enhancing their reliability in real-world deployment. She introduces Tulu 3, a family of fully open post-trained language models. While post-training techniques are essential for refining behaviors and unlocking new capabilities in language models, open approaches have significantly lagged behind proprietary ones. Tulu 3 addresses this gap by providing complete transparency into data, code, and training methodologies, yielding models that outperform comparable open-weight alternatives while narrowing the gap with proprietary systems.