Multimodal models integrate visual, textual, and other data to achieve human-like understanding, but this fusion creates a conflict between cross-modal alignment and modality-specific expertise.The pursuit of unified feature spaces often undermines specialized knowledge in individual modalities, as shown by performance drops in unimodal tasks. To resolve this contradiction, we propose VAMP (Variational Alignment with Modality Preservation), a novel multimodal framework featuring a Dynamic Feature Diversion mechanism that partitions modal representations into two components-one preserving modality-specific expertise and the other enabling cross-modal alignment. Inspired by Variational Canonical Correlation Analysis, we introduce a shared space projection layer that maps features into a common representational space while preserving modality-specific characteristics. We further implement a Progressive Training Strategy that sequentially freezes different components before full fine-tuning, preventing mode collapse and enhancing generalization capabilities. Experimental results demonstrate VAMP's significant performance improvements across zero-shot image classification, cross-modal retrieval, and visual question answering, while simultaneously outperforming baseline models on unimodal tasks. This research provides an engineered solution to the ''knowledge dilution'' problem in cross-modal alignment.

The Uniform Resource Locator (URL) is a primary vector for numerous security threats, including phishing, malware propagation, and spam attacks, making URL-based analysis a critical task in security systems. However, existing research often focuses on static lexical features of individual URLs, overlooking deeper semantic, structural, and behavioral signals that can indicate malicious intent or evasive patterns. In this paper, we propose Behavior-Enhanced Semantic URL Embedding, a novel framework that integrates semantic, structural, and contextual information to improve the detection of security threats embedded in URLs. Our model is composed of three core modules: a semantic understanding module to extract token-level and contextual semantics, a topology structure learning module to capture hierarchical and sequential patterns of URL components, and a downstream multi-task adaptation module that fine-tunes embeddings with supervised contrastive learning for various security detection tasks. We evaluate our method across five public datasets covering key security applications such as malicious URL detection, phishing website identification, and spam filtering, consistently achieving superior performance over existing baselines. Additionally, we demonstrate the extensibility of our approach to related security tasks, showcasing its potential integration into real-world threat detection and security monitoring systems.