MIT News reported on September 14, 2026, that researchers at MIT had developed HardFlow, a method for making generative-model outputs satisfy hard constraints. The underlying arXiv record identifies the paper as “HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization,” by Zeyang Li, Kaveh Alim and Navid Azizan. It was submitted on November 11, 2025, and revised on April 26, 2026.
The paper formulates hard-constrained sampling as a trajectory-optimization problem. The approach is designed to enforce constraints at the terminal output while allowing additional objectives related to sample quality and distribution shift.
MIT News attributed to the researchers reported experiments in robotics, physical-process control and computer vision. In that account, HardFlow satisfied the required constraints while finding better-quality solutions than existing techniques. MIT News also said the method could be applied at deployment time to pretrained generative models without retraining them. Those experimental and deployment claims remain claims reported by an author-affiliated institutional source; the reviewed record does not contain independent validation of the results.
An MIT laboratory page for senior author Navid Azizan lists HardFlow as a 2026 publication in IEEE Transactions on Pattern Analysis and Machine Intelligence. The reviewed material does not establish a journal issue, DOI or peer-review history from that listing, and the listing does not independently establish the method’s experimental validity.
Model Current assessed the reported results as supporting conclusions about the settings tested in the paper, but said they do not establish that HardFlow makes deployed AI systems safe or guarantees constraint satisfaction in arbitrary real-world conditions. The reviewed record therefore does not establish performance across every robot, workspace, unpredictable interaction or deployment environment.
Model Current, an independent analysis published on September 15, said it had not rerun the HardFlow benchmarks and that broader real-world testing remained outstanding. Its assessment treated the reported findings as limited to the tested settings rather than as a general safety guarantee.