New Sampling Framework Boosts Online Feedback-Driven Search

In scientific and engineering fields, maximizing discovery within a limited sampling budget often requires intelligent, observation-guided exploration. While existing generative models are good for local searches, they struggle when preferences are initially unknown and revealed only through sequential feedback, necessitating broader exploration to find high-utility regions. This research addresses this challenge by presenting Bootstrap Flow-Map-Tree (BFMT). BFMT is a novel and computationally efficient sampling framework tailored for global search and alignment, particularly when sampling budgets are tight and historical data is crucial. A key innovation is its ability to construct full tree-paths from any depth using just a single function evaluation, significantly reducing computational overhead and offering critical foresight for subsequent sampling decisions. This allows BFMT to dynamically schedule transition time steps, efficiently allocating its budget to smoothly shift from wide global exploration to precise local refinement of promising areas discovered during the search. Extensive experiments confirm BFMT's superior performance across various search and alignment tasks compared to baseline methods.