HaloProbe: Bayesian Detection and Mitigation of Object Hallucinations in Vision-Language Models

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.

When given an image and a text prompt, vision-language models (VLMs) can produce seemingly sophisticated text responses. However, they are also prone to hallucinate objects in their output, which can have detrimental consequences in safety-critical applications. To detect these hallucinations, certain methods interestingly leverage internal signals from the VLM as it generates the text response; potentially, these methods use the same signals that led to a hallucination to detect it. We conduct an in-depth analysis to understand the robustness of these detection methods. First, we show that a naive baseline of a generated word's token position in the VLM text response achieves competitive performance. Further, we reveal how some detection methods rely on token position and use this heuristic as a shortcut. Second, we explore detection methods' sensitivity to language and vision modalities. Our experiments suggest that strong detection performance is often driven by detecting language-based hallucinations, and when methods rely on token position, they tend to be less visually grounded and exhibit reduced robustness when hallucinations occur earlier in a text response. Lastly, we demonstrate that different detection methods can carry complementary signals; when ensembling methods together, adding weaker methods to an ensemble of higher performing ones still offers performance improvements. Altogether, while there may be useful internal VLM signals to detect hallucinations, our investigation reveals important limitations of the current methods that try to leverage them.

Large Vision-Language Models (LVLMs) have achieved impressive progress in multimodal reasoning, yet they remain prone to object hallucinations, generating descriptions of objects that are not present in the input image. Recent approaches attempt to mitigate hallucinations by suppressing unreliable visual signals in the vision encoder, but many rely on iterative optimization for each input, resulting in substantial inference latency. In this work, we investigate the internal attention dynamics of vision encoders in LVLMs and identify a consistent three-phase structure of visual information processing: diffusion, focus, and rediffusion. Our analysis reveals that hallucination behavior is particularly sensitive to tokens receiving low attention during the focus phase. Motivated by this observation, we propose a lightweight inference-time intervention that selectively suppresses such tokens during the focus phase. The method operates in a training-free manner using statistics from a single forward pass and employs a Determinantal Point Process (DPP) to preserve diverse visual cues while filtering redundant tokens. Extensive experiments across multiple LVLM backbones and decoding strategies demonstrate that the proposed approach consistently reduces hallucination metrics while maintaining competitive caption quality. Moreover, compared to adversarial uncertainty estimation methods, our approach achieves comparable hallucination mitigation with negligible additional inference latency.