Search papers, labs, and topics across Lattice.
University of Electronic Science and Technology
3
0
4
Unlearning sensitive information in LLMs can be achieved without sacrificing performance, with ADU preserving up to 98% of model utility while effectively forgetting.
SPRI achieves a remarkable 3.39 BLEU point improvement over the best existing MoE upcycling method, demonstrating that pretrained weight structures can be effectively leveraged for better expert diversity.
Achieve >95% forget quality in LLMs with minimal side effects by isolating and unlearning tokens within target subdomains using asymmetric LoRA.
