Search papers, labs, and topics across Lattice.
1
0
2
4
Forget likelihood scores and fine-tuning tricks: LLMs leak pre-training data membership through subtle, yet detectable, deviations in their gradient updates.
