A new paper introduces a sampling framework that could make AI-guided scientific discovery significantly cheaper to run.
Researchers propose Bootstrap Flow-Map Tree, or BFMT, a method for guiding generative models through large search spaces when the goal isn't known upfront and can only be learned through sequential feedback. The core trick: BFMT constructs full tree paths from any depth using a single function evaluation, rather than the repeated evaluations that competing methods require. It also schedules when to explore broadly versus when to drill down into promising regions — shifting strategy as evidence accumulates.
This matters because most existing approaches to reward-aligned sampling work well only when they already have a rough idea of where the good solutions are. BFMT is designed for the harder case: open-ended search where high-utility regions must be discovered from scratch. That profile fits drug discovery, materials science, and protein design — domains where compute budgets are finite and promising candidates are rare.
The paper reports that BFMT substantially outperforms baselines across search and alignment tasks, though independent benchmarking on real-world pipelines will be the actual test. For now, it joins a crowded field of inference-time search methods all promising to do more with less.