Tag
The paper introduces Errorquake-10k, a benchmark for evaluating error severity in open-weight LLMs, showing that models with matched accuracy can have vastly different error severity distributions, and argues that severity should be reported alongside accuracy.