Jiantao Zhou

Forbidden capabilities can still be learned even when suppressed in current outputs, as internal gradients enable future acquisition, but gradient sealing effectively blocks these pathways.

Latent LLM personas persist in weights post-training, readily resurfacing with specific prompts, and standard unlearning methods are insufficient to remove them without performance degradation.

CASCADE redefines backdoor detection in multimodal learning by achieving near-perfect accuracy while effectively managing the ambiguity of poisoned samples.

T3HG-Editor achieves unprecedented garment consistency in 3D human editing by embedding body priors from the SMPL-X model, setting a new standard for fidelity in text-driven applications.

DVANet achieves superior performance in image restoration by effectively integrating degradation awareness with advanced visual priors, setting a new standard for handling complex image degradations.