Search papers, labs, and topics across Lattice.
3
58
6
10
LLM agents in high-stakes domains can be verified more reliably by accumulating evidence grounded in expert guidelines, achieving a 12% AUROC improvement and 50% Brier score reduction over existing methods.
LLMs can move beyond simple refusals to actively guide vulnerable users towards safe outcomes, achieving state-of-the-art safety and robustness against jailbreaks.
RealSafe-R1 achieves safety alignment of DeepSeek-R1 without sacrificing reasoning performance, a common trade-off in prior safety alignment efforts.
