Research projects

Selected project pages

A collection of project pages accompanying research on language-model reliability, multimodal representation learning, and model interpretability.

SAMark

A self-anchored semantic watermarking framework for paragraph-level paraphrase robustness.

Distortion-free semantic-level watermarking through constrained sampling over multiple semantic channels.

Autoregressive multi-vector generation in latent space for efficient visual-document retrieval.

Modality-aware multimodal unlearning that separates visual forgetting from textual retention.

Neuron-level analysis of domain-specific mechanisms in multimodal large language models.