output KL divergence

LLM updates can cause regressions; new research explores predictive signals

A new research paper investigates methods to predict when updates to large language models (LLMs) might cause regressions, where a previously correct output becomes incorrect. The study compares various signals, includi…