Speaker: Dan Alistarh (Institute of Science and Technology, Austria – ISTA)

Abstract: Modern large language models have billions to trillions 
of parameters, creating enormous computational and 
memory costs. Quantization, i.e. reducing their 
numerical precision, is the leading practical mitigation 
strategy. But how far can we push it, and what do we 
lose? This talk addresses different sides of this question. 
First, for post-training quantization, we characterize the 
accuracy–compression frontier focusing on large-scale 
evaluations and new formats. Second, for quantizationaware

training, we show that convergence behavior is 
predicted by representation scaling laws, and outline a 
new-state-of-the-art low-precision training method. 
Together, these results offer a statistical framework for 
understanding the interplay between low-precision 
representations and LLM performance.

Biography: Dan Alistarh is a Professor at IST Austria, in Vienna. 
Previously, he was a Staff Researcher with Microsoft, a 
Postdoc at MIT CSAIL, and received his PhD from the 
EPFL. His research is on algorithms for efficient machine 
learning and high-performance computing, with a focus 
on scalable inference and training. In his spare time, he 
works with the ML research team at Neural Magic/Red 
Hat AI on making model compression faster, more 
accurate and broadly accessible.