Search papers, labs, and topics across Lattice.
1
0
3
LLM watermarks, meant to prove AI authorship, can be silently amplified by attackers who compromise the pseudo-random number generator, making models appear to generate more watermarked text than they actually do.
