Xiao Wang

A replay-free GNN architecture effectively combats recency bias in class-incremental learning by routing nodes to experts via masked autoencoder reconstruction error, outperforming existing methods without task IDs.

Spectral disentanglement and knowledge-guided aggregation unlock efficient, high-performance MoE neural operators for large-scale PDE pretraining, slashing expert count by 50% without sacrificing accuracy.

Off-the-shelf vision foundation models fail to transfer dense representations to scientific imagery, but scaling in-domain self-supervised pretraining across 10 million tiles delivers an immediate 7-point Dice improvement under severe annotation limits.