Search papers, labs, and topics across Lattice.
1
0
3
LLMs leak training data membership through subtle, gradient-induced shifts in their internal representations, enabling a surprisingly effective new attack.
