HALP: Detecting Hallucinations in Vision-Language Models without Generating a Single Token

TL;DR

This paper presents a method to predict hallucination risks in vision-language models before text generation by probing internal representations, enabling safer and more efficient model deployment.

Contribution

It introduces a novel approach to detect hallucinations pre-generation using internal model probes across diverse VLMs, outperforming existing post-generation detection methods.

Findings

Probes achieve up to 0.93 AUROC in hallucination detection without decoding.

Pre-generation detection is effective across multiple architectures and modalities.

Different layers and features are most informative depending on the model architecture.

Abstract

Hallucinations remain a persistent challenge for vision-language models (VLMs), which often describe nonexistent objects or fabricate facts. Existing detection methods typically operate after text generation, making intervention both costly and untimely. We investigate whether hallucination risk can instead be predicted before any token is generated by probing a model's internal representations in a single forward pass. Across a diverse set of vision-language tasks and eight modern VLMs, including Llama-3.2-Vision, Gemma-3, Phi-4-VL, and Qwen2.5-VL, we examine three families of internal representations: (i) visual-only features without multimodal fusion, (ii) vision-token representations within the text decoder, and (iii) query-token representations that integrate visual and textual information before generation. Probes trained on these representations achieve strong hallucination-detection performance without decoding, reaching up to 0.93 AUROC on Gemma-3-12B, Phi-4-VL 5.6B, and Molmo 7B. Late query-token states are the most predictive for most models, while visual or mid-layer features dominate in a few architectures (e.g., ~0.79 AUROC for Qwen2.5-VL-7B using visual-only features). These results demonstrate that (1) hallucination risk is detectable pre-generation, (2) the most informative layer and modality vary across architectures, and (3) lightweight probes have the potential to enable early abstention, selective routing, and adaptive decoding to improve both safety and efficiency.