MIT develops HardFlow for hard-constrained generative AI
MIT introduces HardFlow for safety-critical generative AI
MIT researchers have developed HardFlow, an algorithm designed to help pretrained generative AI models meet nonnegotiable requirements in high-stakes settings without retraining. In experiments involving robotic manipulation, maze navigation and text-guided image editing, the method achieved perfect constraint satisfaction while consistently surpassing baseline approaches on solution-quality measures.
The work, published in IEEE Transactions on Pattern Analysis and Machine Intelligence, was led by Zeyang Li with Kaveh Alim and senior author Navid Azizan. The team says the plug-and-play method operates at deployment time and is intended for cases in which safety rules, physical laws or task-specific conditions cannot be violated.
Final outputs matter more than every intermediate sample
Generative models such as diffusion models including Stable Diffusion and flow-matching models such as FLUX create data by transforming random noise. They can generate answers that are close to a request, but safety-critical work may require an output that satisfies every hard condition. A robot route that is nearly valid, for example, can still create a collision risk on a factory floor.
Existing projection-based sampling methods repeatedly impose strict requirements on intermediate samples during generation. MIT’s researchers argue that this can restrict the model’s ability to reach a better final answer. Those methods also commonly emphasize feasibility alone rather than optimizing other properties, such as the distance or duration of a robot trajectory.
Trajectory optimization steers the sampling process
HardFlow instead gives the model freedom to explore during sampling and directs it so that the final output obeys the required constraints. It reframes hard-constrained sampling as a trajectory-optimization task, drawing on optimal-control tools to make incremental corrections along the model’s sampling trajectory.
To make that optimization practical around large neural networks, the researchers used the structure of flow-matching models to split the task into smaller single-step subproblems. Systematic transformations and approximations produced an efficient, scalable algorithm that retains the relevant properties of the original problem while finding a feasible solution at deployment time.
Quality objectives can be handled with safety requirements
The formulation also enables additional objectives alongside hard constraints. In one example, HardFlow can seek a collision-free robot path that is also the shortest route to the target. In robotic-manipulation tests, it avoided obstacles while finding the quickest route to an object; the researchers report that most alternative methods either collided or produced substantially slower paths.
HardFlow’s computation time was comparable to or lower than that of most competing methods. The researchers may later extend the framework so the AI model itself can be updated, allowing constraint satisfaction and sample quality to improve more adaptively. For businesses deploying generative models in controlled physical or regulated workflows, the practical implication is to evaluate whether final outputs meet every mandatory condition while preserving the quality needed for the task.