MIT's HardFlow Keeps AI From Breaking Safety Rules Without Retraining It

MIT researchers have developed a technique that helps generative AI models solve high-stakes problems where a merely plausible answer is not good enough — the output must also satisfy strict, nonnegotiable safety or physical requirements known as "hard constraints." The method, called HardFlow, works with pretrained models without retraining them, and is described in a paper in IEEE Transactions on Pattern Analysis and Machine Intelligence.

Pretrained generative models such as diffusion models like Stable Diffusion and flow-matching models like FLUX create new data by transforming random noise, and excel at producing answers close to satisfying a query. But in safety-critical settings — such as planning a robot's path across a crowded factory floor — a "nearly correct" answer is not good enough: an almost-right path could still let the robot collide with a human co-worker.

The common approach, called projection-based sampling, repeatedly forces the model's partial, intermediate solutions to satisfy strict requirements throughout generation, which can prevent it from reaching a better final solution and often ignores other qualities worth improving, such as a shorter robot path. HardFlow instead gives the model more freedom during generation and enforces hard constraints only on the final output, reformulating the task as a trajectory-optimization problem borrowed from control theory. To make that computation tractable for a huge neural network, the researchers decomposed it into a sequence of smaller, single-step subproblems.

"The promise of generative AI is its ability to explore a rich space of possibilities, but the real world places boundaries on which possibilities are acceptable," said senior author Navid Azizan of MIT's Department of Mechanical Engineering. He worked on the paper with lead author Zeyang Li and co-author Kaveh Alim.

In experiments spanning robotic manipulation, maze navigation and text-guided image editing, HardFlow achieved perfect constraint satisfaction while consistently outperforming baseline methods on solution quality, and its computation time was comparable to or lower than competing methods. In one test, it let a robotic manipulator avoid collisions with obstacles while finding the quickest path to its target, while most other methods either caused collisions or took significantly longer.