About the Tech Talk
Join Scale AI researchers as they present their NeurIPS 2024 main track paper, Learning Goal-Conditioned Representations for Language Reward Models, introducing a novel approach to improving LLM alignment. This technical session will explore how goal-conditioned representations can enhance reward modeling and significantly reduce computational costs. Through a detailed examination of the methodology and results, we'll demonstrate how this approach achieves substantial improvements in both model performance and decoding efficiency. The presentation will be followed by an in-depth discussion of practical implementation considerations and future research directions.
Key takeaways
- Discover a novel architecture for reward modeling that achieves large improvements on both math and natural language tasks.
- Learn how to implement goal-conditioned rewards into post-training and decoding pipelines to filter 55% of tokens without accuracy loss, significantly reducing computational costs.
- Explore how this research advances LLM alignment and control, setting new directions for developing more reliable, efficient, and interpretable AI systems.
About Sean Hendryx | Speaker
Sean Hendryx leads frontier data research and applied ML for generative AI at Scale. He has been developing models for collaborative human-AI systems for over six years. He is currently focused on researching systems that most effectively leverage human supervision for the advancement of frontier models and has publications on alignment, frontier evaluations, test time compute scaling, meta-learning, and online learning.
About Vaskar Nath | Speaker
Vaskar Nath is a Machine Learning Research Engineer at Scale AI, where he works on advancing the capabilities of language models. His research focuses on improving reward modeling, enhancing tool usage, and strengthening reasoning abilities. Vaskar holds a bachelor’s degree from the University of Toronto, where he double-majored in computer science and mathematics with a minor in statistics. Prior to joining Scale, he interned at Meta, Nuro, AWS, and Intel.
About Jessica Zhu | Moderator
Jessica Zhu is the tech lead of Scale's Generative AI Growth team. She has spent the past few years working on various products at the forefront of AI, including document processing, computer vision, and LLM infrastructure. In her free time she enjoys traveling, boxing, and playing board games.
