Sheng Li

By ditching GNN inductive biases, MANDATE achieves superior graph fraud detection using a Transformer architecture with multi-scale positional encoding and adaptive handling of homophily.

Achieve state-of-the-art arbitrary-scale super-resolution with a single model by reframing extreme upscaling as a series of distribution-aligned, self-similarity-aware transitions.

LLMs can be made significantly more robust to jailbreaks by weighting the reasoning steps in DPO training, leading to more principled refusals.