Using a large language model to synthesize the initial candidate solutions ('warm starts') that seed an active learner for multi-objective optimization, so fewer expensive labels are needed to find good configurations. Students learn how warm starts change an active learner's sample efficiency, how LLM-proposed guesses compare with Gaussian Process models and Tree of Parzen Estimators as baselines, and the empirical finding that LLM warm starts help most on low- and medium-dimensional problems while Bayesian surrogate methods dominate in high dimensions.
An empirical study of whether Large Language Models can supply the initial guesses ('warm starts') for active learning on multi-objective software-engineering optimization tasks. Active learners labe…