Advancing Grounded Multimodal NER via Self-reflective Prompt Refinement and Visual Noise Mitigation

Grounded Multimodal Named Entity Recognition (GMNER) aims to jointly identify named entities, their semantic types, and corresponding locations in images from text-image pairs. While Multimodal Large Language Models (MLLMs) show great potential for GMNER due to their ability to process visual and textual information simultaneously, their application faces two critical challenges: the difficulty of defining effective task-specific prompts for specialized domains and the visual noise interference that degrades recognition performance. To address these challenges, we propose Self-Reflective Prompt Refinement (SRPR) and a Multi-Model NER-VG Framework. SRPR leverages reasoning models to automatically optimize prompts through iterative self-reflection, while the multi-model framework separates textual entity recognition from visual grounding to mitigate visual noise interference. Our approach achieved first place in the CCKS 2025 evaluation with an F1 score of 65.78%, demonstrating its effectiveness for domain-specific applications.