cs.AI, cs.CL, cs.CR

Layerwise Convergence Fingerprints for Runtime Misbehavior Detection in Large Language Models

arXiv:2604.24542v1 Announce Type: cross
Abstract: Large language models deployed at runtime can misbehave in ways that clean-data validation cannot anticipate: training-time backdoors lie dormant until triggered, jailbreaks subvert safety alignment, a…