Search papers, labs, and topics across Lattice.
SB Intuitions Corp
3
0
4
Text safety neurons in VLMs are localized and critical, while visual safety mechanisms are diffuse and complex, revealing a fundamental gap in current alignment strategies.
Robust CLIP models, surprisingly, can be *more* vulnerable to natural semantic variations than standard CLIP, revealing a critical flaw in current robustness training strategies.
Stop reacting to unsafe LLM outputs and start predicting them: StreamGuard forecasts the harmfulness of future text, enabling earlier and more effective safety interventions.
