Why Jailbreaks Succeed in Diffusion Language Models: An Energy Landscape Analysis
Three complementary, training-free detection signals are derived that reads the initial safety disposition from the logit distribution before generation begins, and two trajectory-velocity signals that track kinetic energy in complementary subspaces of the logit space, so the three signals cover each other's blind spot...
Thong Bach, Dung Nguyen, T. Le et al.
· 0 citations