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.
Abstract (translated)
幻觉仍然是视觉-语言模型（VLMs）面临的一个持久挑战，这些模型常常描述不存在的对象或捏造事实。现有的检测方法通常在文本生成之后操作，这使得干预既费时又不够及时。我们研究了是否可以在任何标记生成之前通过探测模型的内部表示来预测幻觉风险，在一次前向传递中完成这一任务。我们对一系列视觉-语言任务和八种现代VLM进行了调查，包括Llama-3.2-Vision、Gemma-3、Phi-4-VL 和 Qwen2.5-VL。我们研究了三种内部表示类型：(i) 仅视觉特征而不进行多模态融合；(ii) 文本解码器内的视觉标记表示；以及 (iii) 在生成前整合视觉和文本信息的查询标记表示。基于这些表示训练的探测器在不进行解码的情况下达到了很强的幻觉检测性能，例如，在Gemma-3-12B、Phi-4-VL 5.6B 和 Molmo 7B 上实现了高达0.93的AUROC（受试者工作特征曲线下的面积）。对于大多数模型来说，晚期查询标记状态最为预测性，而视觉或中间层特性在某些架构中占主导地位（例如，在使用仅视觉特征时，Qwen2.5-VL-7B 实现了约0.79的AUROC）。 这些结果表明：(1) 幻觉风险可以在生成之前被检测到；(2) 最有信息量的层和模态在不同架构之间有所不同；以及 (3) 轻量级探测器具有潜力，能够实现早期放弃、选择性路由和自适应解码，以提高模型的安全性和效率。
URL
https://arxiv.org/abs/2603.05465
