Search papers, labs, and topics across Lattice.
1
0
3
LLM safety probes can be made significantly more robust to adversarial attacks by requiring consistent evidence across token segments, not just isolated spikes.
