GISTBench: Evaluating LLM User Understanding via Evidence-Based Interest Verification
GISTBench creates a benchmark to test if Large Language Models (LLMs) can accurately understand user interests from their interaction history in recommendation systems. It introduces novel metrics, Interest Groundedness (IG) and Interest Specificity (IS), to verify whether predicted interests are actually supported by behavioral evidence, moving beyond simple accuracy scores.