Search papers, labs, and topics across Lattice.
1
0
3
0
LLMs leak uniquely identifiable fingerprints in their attention patterns when processing semantically conflicting prompts, enabling robust provenance verification even after fine-tuning, pruning, or merging.
