Sampling Meets Interaction: Sequentially-Controlled Multi-Particle Flow-Maps for Efficient Inference-time Search
IMPFM is a framework for sample-efficient online search that uses multiple particles to explore a target distribution. It combines flow maps for collective posterior sharing with an interaction-aware Feynman–Kac corrector. This allows particles to adapt quickly to feedback while maintaining broad coverage, leading to higher reward scores with fewer interactions and resisting reward over-optimization.