Id: arxiv/2607.04513 Url: https://arxiv.org/abs/2607.04513 Tags: math.OC, cs.LG Score: 0 Title: Constrained Flow Matching via Lagrangian Dual Flows Domain: math.OC Source: arxiv Authors: Vince Kurtz, Alexander Davydov Reviews: 0 Subject: math.OC Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-07-05 Page: https://delphi.science/r/arxiv/2607.04513 Flow matching is a powerful tool for generative modeling, but emerging applications in robotics, planning, and control require inference-time constraints on generated outputs. Such constraints are often complex and highly nonlinear. As a result, methods designed for linear constraints like image inpainting are rarely sufficient, and projection or optimization-based alternatives can be prohibitively expensive. In this paper, we introduce Lagrangian Dual Flows, a new family of constrained generation techniques based on Lagrangian dual dynamics. By flowing a dual co-state alongside generated…