HalDec-Bench: Benchmarking Hallucination Detector in Image Captioning

Abstract

Hallucination detection in captions (HalDec) assesses a vision-language model's ability to correctly align image content with text by identifying errors in captions that misrepresent the image. Beyond evaluation, effective hallucination detection is also essential for curating high-quality image-caption pairs used to train VLMs. However, the generalizability of VLMs as hallucination detectors across different captioning models and hallucination types remains unclear due to the lack of a comprehensive benchmark. In this work, we introduce HalDec-Bench, a benchmark designed to evaluate hallucination detectors in a principled and interpretable manner. HalDec-Bench contains captions generated by diverse VLMs together with human annotations indicating the presence of hallucinations, detailed hallucination-type categories, and segment-level labels. The benchmark provides tasks with a wide range of difficulty levels and reveals performance differences across models that are not visible in existing multimodal reasoning or alignment benchmarks. Our analysis further uncovers two key findings. First, detectors tend to recognize sentences appearing at the beginning of a response as correct, regardless of their actual correctness. Second, our experiments suggest that dataset noise can be substantially reduced by using strong VLMs as filters while employing recent VLMs as caption generators. Our project page is available at https://dahlian00.github.io/HalDec-Bench-Page/.

Figures (27)

• Figure 1: We introduce a novel benchmark, HalDec-Bench, which evaluates the VLM's ability to detect hallucinations in captions. We employ state-of-the-art VLMs to generate image-caption pairs (Captioner) and manually annotate the hallucinated parts when present. We benchmark diverse VLMs as Detector and find that subtle hallucinations can be hard to detect, even by state-of-the-art VLMs.
• Figure 2: Hallucinated sentences in HalDec-Bench. Hallucinated portions are often subtle, requiring fine-grained image-text alignment ability to detect them.
• Figure 3: Details of our HalDec-Bench. (a) We confirm that our dataset includes diverse caption length. (b) We also annotate hallucination types into 8 categories. The rarest type is Text, but it still includes nearly 500 instances, enabling deeper analysis. (c) We create captions using diverse VLMs as Captioners, enabling analysis with combinations of diverse Detector models. (d) An example of an annotation.
• Figure 4: Examples of incorrect sentences with detectors’ correctness scores. Higher scores indicate greater confidence in correctness. Detectors are prone to being overconfident in these examples. We highlight detectors’ errors in red within the text and mark the grounded incorrect regions in the image with orange boxes.
• Figure 5: The size of plots indicates the parameter size. Left: MMMU performance measured on Captioners (X-axis) vs. AUROC measured by GPT-5-mini (Y-axis) for each Captioner. Advanced Captioners tend to produce errors that are difficult to detect. Right: MMMU (X-axis) vs. AUROC (Y-axis) for each detector. Detectors with better MMMU performance do not necessarily perform better on HalDec-Bench.