HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization

HardFlow is a new method for controlling AI generative models (systems that create new data) to satisfy hard constraints, which are strict requirements that must be met without exception. Instead of forcing the entire generation process to stay within allowed boundaries, HardFlow reformulates the problem as trajectory optimization (finding the best path through a decision space), using techniques from control theory to ensure constraints are satisfied at the end of generation while maintaining sample quality.