ASAP Seminar Series

About The Seminar

The ASAP is a fully virtual seminar that, while emphasizing efficiency as its name suggests, takes a distinct approach from traditional ML systems seminars. As foundation models grow increasingly powerful through scaling of data and model size, their capabilities remain inherently limited by their architectures. This seminar explores novel architectural designs and paradigms for both training and inference, examining sequence modeling approaches from an algorithmic perspective. The seminar serves as a bridge between theoretical, algorithmic, and systems research communities, tackling fundamental challenges in Transformer models—including reasoning capabilities, generalization power, memorization, and expressiveness. By adopting first-principles approaches, it seeks solutions that address these limitations while maintaining computational efficiency on modern hardware, with the ultimate goal of motivating the next generation of foundation models that overcome current limitations.