Wednesday / Student Seminar

April 1 / 12:00 PM

Mudd 3514

Speaker: Robin Luo

Title: Softmax1 Is All You Need

Abstract: Foundation models have achieved strong performance across a wide range of domains, but their growing scale also raises important challenges in efficiency, robustness, and safety. In this talk, I will present recent work centered on Softmax1, a simple yet effective modification that improves foundation models by enabling more efficient reasoning, reducing outlier behavior, and enhancing robustness under quantization and safety alignment settings. I will discuss both the core intuition behind this approach and its broader implications for building trustworthy and efficient foundation models.

Bio: Robin (Haozheng) Luo is a Ph.D. candidate in Computer Science at Northwestern University and a DAAD AINeT Research Fellow and Adobe Research Fellow. His research focuses on trustworthy and efficient foundation models, including reasoning efficiency, quantization-robust architectures, and safety alignment, with publications at top-tier conferences such as NeurIPS, ICML, ICLR, and USENIX Security.