Is In-Domain Training Enough for Fine-Grained Industrial Anomaly Understanding?
The study investigates whether fine-tuning a single multimodal large language model (MLLM) is sufficient for industrial anomaly understanding, which requires detection, localization, description, and reasoning. Findings show that specialization leaves models uneven across tasks. The proposed SiGMA framework improves performance by combining heterogeneous frozen MLLMs with a dedicated visual defect expert to boost accuracy.