Search papers, labs, and topics across Lattice.
1
0
5
32
Continuously nudging LLM activations during generation can effectively correct misalignment without sacrificing coherence, offering a lightweight runtime defense against adversarial prompts and other triggers.
