What AI says about Hal-Eval

Hal-Eval is a universal, fine-grained evaluation framework for large vision-language models (LVLMs) designed to assess hallucinations in model outputs. It introduces a refined taxonomy that includes Event Hallucination and uses large language models to generate and filter fine-grained hallucination data. As a benchmark, Hal-Eval enables evaluation across a broad spectrum of hallucinations and supports both discriminative and generative assessment, with a demo (Hal-Evaluator) and an accompanying preprint paper.