Search papers, labs, and topics across Lattice.
Affiliation:
1
0
3
Deeper reasoning creates an unexpected safety backdoor: extended CoT dilutes attention away from system constraints, making models progressively more vulnerable to jailbreaks the longer they "think."
