Jiahao Xu

Generating realistic and stable human co-manipulation motions is now possible by explicitly modeling object affordances and spatial configurations within a flow-matching framework.

Existing LLM watermarking schemes crumble when text is short, but XMark maintains high decoding accuracy and text quality even with limited tokens.

LLMs can boost few-shot learning for pathology images, but only if you dynamically adapt the language priors to each image and stochastically integrate multiple "expert" descriptions.

Heterogeneous federated LLM fine-tuning gets a boost from parallel one-rank adaptation, sidestepping the noise issues that plague existing LoRA-based methods.