HalLoc: Token-level Localization of Hallucinations for Large Vision Language Models

Hallucinations present a significant challenge to the reliability of large vision-language models. Identifying hallucinated segments within model outputs is a natural strategy for addressing this issue. Precise identification of hallucinated segments can facilitate a better understanding of hallucination patterns, evaluation of the fidelity of the generated outputs, and development of methods to revise them. Despite its importance, hallucination localization has mainly been under-explored, especially in the context of large vision-language models. This work introduces HalLoc-Bench, the first benchmark especially designed for hallucination localization. HalLoc-Bench supports the training and evaluation of localization across various hallucination types (object, attribute, relation, scene) and tasks (VQA, captioning, instruction-following). We also present HalLoc, a simple yet effective localizer that sets a strong baseline on HalLoc-Bench. Experiments show that HalLoc-Bench effectively assesses hallucination localization, providing a valuable tool for advancing this field. Our analysis further reveals that better hallucination localization can improve the evaluation and mitigation of hallucinations in large vision-language models. HalLoc-Bench is available here.