MoLD: Fine-Grained Multimodal Risk Assessment via Dynamic Analysis Weights

Pingyu Wu, Pengcheng Li, Wenbin Liao, Zuolong Li, Wenbo Zhou, Weiming Zhang, Nenghai Yu

Keywords: Large Model, Multimodal, Dynamic Weight Analysis, Few-shot, Cross-scale, Risk Assessment, Interpretability

TL;DR: A novel framework for interpretable and efficient multimodal risk assessment via dynamic analysis of multiple LoRA module weights

Abstract: Current multimodal AI safety detection often lacks granularity, interpretability, and adaptability. To address these limitations, we introduce **MoLD** (Mixture of LoRA Detectors), a framework that uniquely assesses risk by dynamically analyzing the interplay of multiple Low-Rank Adaptation (LoRA) module weights. This approach yields fine-grained, interpretable assessments beyond binary classification, enables concurrent **multi-risk detection**, maintains robustness on long-sequence data, and supports low-cost modularity. Impressively, MoLD demonstrates state-of-the-art (**SOTA**) performance on textual and visual benchmarks while achieving exceptional **few-shot** learning, reducing data requirements by over **90\%**. Thus, MoLD provides a powerful, scalable, and data-efficient path to robust, interpretable risk assessment in large-scale multimodal AI systems.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 2034