Search papers, labs, and topics across Lattice.
Xiamen University
4
0
6
7
Agents often prioritize user alignment over factual accuracy, and MemSyco-Bench reveals just how much memory influences this dangerous trend.
Current VLM unlearning methods mostly just hide the data you want gone, and can be easily bypassed with clever prompting or retraining.
LLMs can forget harmful knowledge with just a few examples, thanks to a new method that surgically overwrites sensitive representations.
LLMs struggle with cross-document relation extraction because of the sheer number of possible relations, but a hierarchical classification approach can unlock their potential.
