Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Affiliation:
4
0
5
1
Repeated cheating in verifiable LLM inference can exploit security gaps, but a new mechanism cuts deviation profits by over 30% without compromising verification speed.
Fine-tuning's "small" updates can hide large, malicious changes, but this work offers cryptographic proofs to ensure updates stay within defined boundaries, regardless of model size.
Widely proposed VDF-based randomness beacons are economically insecure with delays of only a few seconds, as rational adversaries can profit by exploiting MEV opportunities with faster hardware.
Optimistic rollups can now handle regulated finance: RegGuard slashes settlement failures by 90% and eliminates detectable ordering manipulation.