Recent anomaly detection methods built on multimodal large language models (MLLMs) can output textual explanations alongside their predictions, which helps users interpret why an image is flagged as anomalous. But judging whether those explanations are accurate and useful is not straightforward, and existing evaluation approaches may be costly or unreliable.
The new arXiv preprint introduces AnomalyJev, a framework designed to evaluate image anomaly explanations through claim-level verification. Instead of treating an explanation as a single block, AnomalyJev breaks it into individual claims and checks each one, aiming to make the evaluation process more reliable and efficient.
The paper is a preprint, so the details are preliminary, but the core idea is to give researchers a structured way to assess the quality of MLLM-generated anomaly explanations. The source abstract does not provide experimental results or comparisons with other evaluation methods, so the contribution at this stage is the proposed framework itself.