Search papers, labs, and topics across Lattice.
1
0
2
Jailbreaking reasoning models turns out to be an effective antidote to RL-induced overconfidence, recovering suppressed pre-alignment entropy to slash black-box calibration error by up to 5脳.
