Search papers, labs, and topics across Lattice.
2
0
6
4
The propensity of frontier agents to overclaim, a misrepresentation that can mislead the user, is quantified and results show that agents'final responses are not reliable accounts of their actions.
Continuously nudging LLM activations during generation can effectively correct misalignment without sacrificing coherence, offering a lightweight runtime defense against adversarial prompts and other triggers.
