Topics

GOLLuM trains LLMs using Bayesian uncertainty signals, raising high-yield Buchwald-Hartwig discovery from 24% to 43% in 50 experiments across 19 benchmarks.