Zhikun Zhang

Forget-critical experts in MoE models can be under-regularized during unlearning, but TRACE rebalances their activation to significantly boost utility.

PINNs can now jointly infer parameters and detect regime-switching in dynamical systems, outperforming traditional methods that treat these tasks separately.

LLMs can selectively forget specific knowledge without sacrificing overall performance, thanks to a new representation-deviation technique that precisely targets and disrupts unwanted associations.