Detecting Vision-Language Model Hallucinations before Generation

ACL ARR 2025 May Submission7161 Authors

Abstract: Object hallucination is a significant challenge that undermines the reliability of the Vision Language Model (VLM). Current methods for evaluating hallucination often require computationally expensive complete sequence generation, making rapid assessment or large-scale analysis difficult. We introduce HALP (HALlucination Prediction via Probing), a novel framework to efficiently estimate a VLM's propensity to hallucinate objects without requiring full caption generation. HALP trains a lightweight probe on internal VLM representations extracted after image processing but before autoregressive decoding. HALP offers a new paradigm for efficient evaluation of VLM, a better understanding of how VLMs internally represent information related to grounding and hallucination, and the potential for real-time assessment of hallucination risk.

Paper Type: Short

Research Area: Multimodality and Language Grounding to Vision, Robotics and Beyond

Research Area Keywords: Interpretability, Explainability, Hallucination in VLMs, Vision Language Models, Multimodal AI

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 7161