Search papers, labs, and topics across Lattice.
University of North Carolina at Chapel Hill
2
0
6
0
LLMs can now adaptively survey the *right* people, not just ask the *right* questions, boosting response prediction by 12% while querying only 10% of the population.
LLM judges are surprisingly susceptible to subtle rubric manipulations that can induce significant preference drift, even while maintaining benchmark performance, creating a stealthy attack surface for biasing model alignment.
