Tianyu Zhan

The paper introduces ProjLens, an interpretability framework that reveals that backdoor vulnerabilities in Multimodal Large Language Models (MLLMs) are encoded within a low-rank subspace of the projector, causing a measurable semantic shift in poisoned inputs.

A multi-agent framework is proposed to sanitize retrieved content in Retrieval-Augmented Generation while preserving semantic core and reducing privacy leakage.

This paper proposes TSGR, a generative retrieval framework for industrial e-commerce search that incorporates value awareness into item representation and candidate ranking.