ResearchPod Summary
Vision-language models (VLMs) frequently suffer from hallucinations, where generated text does not accurately reflect the visual input. While existing benchmarks often focus on binary detection of these errors, they lack the granularity to explain why a mismatch occurs or to ground the error in specific visual regions. The authors propose GAVEL (Grounded Caption Error Verification and Localization) to address this gap by requiring models to perform joint verification, explanation, and localization.
To support this task, the authors constructed a large-scale dataset containing 35,000 training pairs and 5,000 test pairs, featuring diverse hallucination types (e.g., object, color, spatial, and action errors). The dataset was built using a semi-automatic pipeline that generated candidate image-text pairs, filtered them for low alignment using a VLM, and then utilized human annotators to provide natural language explanations and bounding-box coordinates for the errors. The authors evaluated several state-of-the-art closed-source and open-source models on this benchmark and trained a supervised baseline to assess the learnability of the grounding signals.
The experiments reveal that even high-performing closed-source models struggle with the GAVEL task, highlighting the difficulty of fine-grained hallucination grounding. While GPT-5 showed strong performance in coarse localization, other models exhibited varying degrees of success across different hallucination categories. Crucially, the supervised baseline demonstrated consistent improvements after training on the GAVEL dataset, confirming that the provided annotations offer effective supervision for improving model reliability and interpretability.
By moving beyond simple binary feedback, GAVEL provides a diagnostic framework that helps researchers understand the specific causes of vision-language misalignment. This is essential for developing more reliable generative systems, as it enables automatic error correction and iterative refinement, ultimately leading to more interpretable and trustworthy AI agents.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.