Search papers, labs, and topics across Lattice.
University of California, Berkeley
1
0
3
Retrieval-augmented LLMs are surprisingly vulnerable to memory poisoning via synonym substitution, a loophole that gradient-based defenses can't close.
