Search papers, labs, and topics across Lattice.
University of Padua
2
0
6
LLMs telegraph their refusals internally, letting you cut adversarial prompt search time by up to 72% by listening to their "gut feelings" before they even speak.
Discarded classification weights from pre-trained vision models can be resurrected as powerful semantic prototypes, unlocking zero-shot vision-language alignment and boosting cross-modal performance.
