Search papers, labs, and topics across Lattice.
1
0
3
LLM agents in a simulated NYC learn to selectively trust and deceive, but remain surprisingly vulnerable to adversarial steering, highlighting a fundamental safety-helpfulness trade-off.
