Xiaoli Liu

LLMs' semantic matching accuracy in agentic AI collapses beyond a discrimination capacity crossover, revealing that simply scaling context windows or compressing pipelines won't solve the problem – you need bigger, better models.

Ditching predefined functional sub-networks unlocks state-of-the-art brain disorder diagnosis by learning hierarchical brain network organization directly from fMRI data.

BrainSTR disentangles subtle disease signatures in dynamic brain networks by explicitly modeling spatio-temporal dependencies with contrastive learning, revealing interpretable biomarkers for neuropsychiatric disorders.

Achieve state-of-the-art atmospheric turbulence mitigation by modeling non-isoplanatic blur with Gaussian splatting, outperforming existing methods on both synthetic and real-world datasets.