Search papers, labs, and topics across Lattice.
Carnegie Mellon University
1
0
3
Allowing language models to explore unsafe reasoning can actually enhance their ability to discern harmful from harmless prompts, reducing over-refusal without sacrificing safety.
