Search papers, labs, and topics across Lattice.
1
0
3
7
LLMs leak risk signals in their hidden state trajectories during decoding, enabling a 95% effective jailbreak defense without training or modifying the model.
