A1527
Title: Performance guarantees for LLMs
Authors: Carey Priebe - Johns Hopkins University (United States) [presenting]
Abstract: Performance guarantees for LLMs are essential to myriad critical applications -- enabling any application for which quantifiable confidence in LLM performance is nonnegotiable. The Data Kernel Perspective Space (DKPS) provides a framework to generate empirically validated theoretical predictions allowing guaranteed performance quantification for these transformative models. We will present the probability/statistics/linear algebra for DKPS, and illustrative examples. The probability/statistics/linear algebra for DKPS is analogous to that introduced for network time series recently.
