Search papers, labs, and topics across Lattice.
Affiliation:
2
0
5
Static input guardrails remain blind to silent system prompt overrides, but inspecting post-generation latent trajectories catches constraint-violating responses before delivery while slashing false positive rates to under 3%.
HarmVideoBench reveals that existing benchmarks miss critical layers of harmful video understanding, while a new method boosts model accuracy by over 20%.
