Researchers at MIT have developed an algorithm named HardFlow, designed to ensure that generative AI models satisfy strict, "hard" requirements on their final output. Unlike existing methods that attempt to force compliance at every intermediate step of the generation process, HardFlow treats constraint satisfaction as a control problem. It applies adjustments to the velocity field—the mechanism used in flow-matching techniques to steer generation from randomness to a finished result—so that the constraints are met only at the final stage.

The researchers note that by not requiring every intermediate step to satisfy the constraints, the model maintains more freedom to explore and find high-quality solutions. This approach is particularly relevant for high-stakes or safety-critical applications, such as robotics, physical process control, and computer vision, where following precise rules is essential.

In computer simulations covering tasks such as obstacle avoidance for robotic arms, maze navigation, and image editing, HardFlow met the required constraints every time. The tests showed that the method produced higher-quality results than rival methods while maintaining comparable or lower computational times. Notably, the method can be applied directly to already-trained models without the need for retraining.

Current results are based entirely on simulations, including the D3IL and Maze2D simulators. Independent laboratory reproduction of these results has not yet been reported, and it remains to be seen if the method can handle the types of fuzzy safety rules relevant to large language models or if it extends to other model families beyond flow-matching.

Sources: