MIT unveils HardFlow so generative AI can obey hard safety rules without retraining

MIT researchers published HardFlow, a method that steers already-trained generative AI so the final answer meets strict safety or physics rules — without retraining the model.

When a robot path is almost safe, it can still hit someone. MIT now has a named, peer-reviewed method that tries to keep the last answer inside hard rules — without retraining the model. That is a lab result, not a product on a factory floor.

Adam Zewe, writing for MIT News, reported a new method called HardFlow. The paper is by lead author Zeyang Li and Kaveh Alim, both MIT graduate students, and senior author Navid Azizan, an MIT professor in mechanical engineering and in the Institute for Data, Systems, and Society. The research appears this week in IEEE Transactions on Pattern Analysis and Machine Intelligence — a peer-reviewed computer-vision and machine-learning journal. That MIT News page is the filing event. This desk did not rerun the models.

Hard constraints, as the authors use the phrase, are nonnegotiable safety, physical, or task rules. A “pretty close” answer is not enough. MIT News’s factory-floor example: a robot path that is nearly correct can still hit a human co-worker. File that as the news-office example — not a reported accident.

The usual fix keeps forcing the model’s in-progress drafts to obey the rules at every step. The authors say that can block a better final answer, and that it often only chases the hard rules — missing a chance to also make the answer better, such as a shorter robot path.

HardFlow instead lets the model explore more freely while it is still drafting, then enforces the hard rules on the last output. Azizan, quoted by MIT News: “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. Our approach lets us preserve that generative power while enforcing the nonnegotiable requirements of high-stakes or safety-critical applications.” Lead author Li, quoted on the same page: what matters is the final output, because the internal drafts are thrown away.

The method is plug-and-play at use time. It can sit on already-trained generative models without retraining them. MIT News names common pretrained image models as the kind of system this can attach to. How the sampler is wired is in Sources.

MIT News says experiments spanned robotics, control of physical systems, and computer vision. Across robotic manipulation, maze navigation, and text-guided image editing, the authors say HardFlow met every required constraint and beat existing methods on solution quality. On a 50-trial robot-arm task, they report a perfect safety rate — 1.00, meaning no collisions in those trials — and the fewest steps to the target, about 52.5 on average. Other methods they tested either hit obstacles or took longer. Those figures are the authors’. This desk did not rerun the trials.

The authors also say compute time was comparable to or lower than most competing methods. A later step they flag: they could extend the method so the model itself can be updated. File that as future work — not a shipped product.

CONFIRMED here: the 14 Sep MIT News primary and the IEEE TPAMI HardFlow paper. NOT claimed: a product launch, a factory deployment, FDA clearance, or that HardFlow “solves AI safety.” Distinct from the already-filed nvidia-skild-ai-s1-one-video, amodei-pace-the-frontier, and fda-genai-medical-devices-discussion-paper — this filing is the MIT method paper only.