Ensembling Multiple Hallucination Detectors Trained on VLLM Internal Representations

Abstract

This paper presents the 5th place solution by our team, y3h2, for the Meta CRAG-MM Challenge at KDD Cup 2025. The CRAG-MM benchmark is a visual question answering (VQA) dataset focused on factual questions about images, including egocentric images. The competition was contested based on VQA accuracy, as judged by an LLM-based automatic evaluator. Since incorrect answers result in negative scores, our strategy focused on reducing hallucinations from the internal representations of the VLM. Specifically, we trained logistic regression-based hallucination detection models using both the hidden_state and the outputs of specific attention heads. We then employed an ensemble of these models. As a result, while our method sacrificed some correct answers, it significantly reduced hallucinations and allowed us to place among the top entries on the final leaderboard.