A new arXiv paper introduces ConicFlow, a method for constrained sampling from flow matching models at inference time. The authors frame the problem as an optimal control task, where the goal is to balance staying close to the pretrained trajectory while maximising a reward, all under hard terminal constraints.
This formulation allows constraints to be applied directly during generation, rather than requiring retraining or fine-tuning. The paper's abstract describes the core trade-off explicitly: deviation from the pretrained path is weighed against reward, with the terminal condition enforced as a hard requirement.
Since this is a single source, the summary reflects only the paper's stated contribution. No experimental results or comparisons are available from the abstract alone.