Search papers, labs, and topics across Lattice.
University of Illinois Urbana-Champaign
2
0
5
6
Skill-switching accuracy in LLMs drops significantly on complex tasks, but a new training approach boosts performance from 34.4% to 68.4% on challenging benchmarks.
LLMs fail to deliver personalized responses that align with human judgments, often producing results indistinguishable from generic outputs.
