{"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","summary":"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…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-07-05","reactions":{}}