MIT researchers reported on 14 September 2026 a technique, called HardFlow, that aims to let generative AI models produce outputs meeting nonnegotiable safety, physical or task requirements while preserving output quality. The work appeared in IEEE Transactions on Pattern Analysis and Machine Intelligence, with Navid Azizan as senior author and Zeyang Li as lead author.

The approach gives the model more freedom during generation and enforces the hard constraints only on the final output, rather than at every intermediate step. It is described as plug-and-play and applicable at deployment time to pretrained models without retraining.

In the reported experiments involving robotic manipulation, maze navigation and text-guided image editing, HardFlow achieved perfect constraint satisfaction and outperformed baseline methods on solution quality, while computation time was comparable to or lower than most competitors. The researchers said future work could extend the framework to settings where the model itself can be updated.