ConicFlow: constrained Inference-time optimal control for flow matching models
We propose ConicFlow, which frames constrained sampling from a flow matching model at inference time as an optimal control problem that balances deviation from the pretrained trajectory against reward, subject to hard terminal constraints.
The article introduces ConicFlow, a method that addresses constrained sampling from flow matching models during inference by formulating it as an optimal control problem, balancing deviation from a pretrained trajectory with reward under hard terminal constraints.
What happened
The article presents ConicFlow, a novel approach to constrained sampling in flow matching models. It formulates the problem as an optimal control task, using receding-horizon control and a primal-dual algorithm to optimize short trajectory windows. The terminal-estimate variant improves accuracy by evaluating constraints at the predicted terminal point. The method is validated through experiments in low- and high-dimensional settings.
Why it matters
ConicFlow offers a computationally efficient way to enforce constraints during inference, which is critical for applications requiring safety and reliability. Its ability to balance reward and constraints while minimizing deviation from pretrained models enhances the practicality of flow matching in real-world scenarios.
- ConicFlow provides a structured approach to constrained inference in flow models
- The receding-horizon method enhances computational feasibility
- The terminal-estimate variant improves accuracy and stability
The bigger picture
The method improves the reliability and efficiency of flow matching models in constrained environments, which could lead to better performance in applications like robotics, autonomous systems, and safety-critical AI deployments.
ConicFlow introduces an effective method for constrained sampling in flow matching models, demonstrating strong performance in both low- and high-dimensional settings while maintaining alignment with pretrained trajectories.
Technical details
- Optimal control problem formulation for constrained sampling
- Receding-horizon control with short trajectory window optimization
- Receding-horizon primal-dual algorithm for subproblem resolution
- Terminal-estimate variant to avoid local truncation errors
- Convex-affine structure analysis for suboptimality bounds
FAQ
What is ConicFlow?
ConicFlow is a method that frames constrained sampling from flow matching models as an optimal control problem, balancing deviation from a pretrained trajectory with reward under hard terminal constraints.
How does ConicFlow handle computational tractability?
It uses receding-horizon control, optimizing only a short window of the trajectory at each step, which makes the problem computationally feasible.
What are the benefits of the terminal-estimate variant?
The terminal-estimate variant avoids accumulating local truncation errors by evaluating the reward and constraints at the predicted terminal point rather than the end of the current horizon.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv - cs.AI on Oct 14, 2026.
Comments
Sign in to join the discussion.