Letting an AI break the rules in the middle, as long as the answer is right at the end

Most generative AI is good at "close enough." That's fine for a picture or a paragraph. It's useless for a robot arm that must not hit the wall. Some requirements aren't preferences. They're hard constraints, and an answer that misses by a little is just wrong.

A group at MIT published a method for this last week, called HardFlow. The paper is in IEEE Transactions on Pattern Analysis and Machine Intelligence, with Zeyang Li as lead author and Navid Azizan as senior author.

What I liked is how simple the core idea is. These models build an output in many small steps. The usual fix is to force the constraint at every single step, so the answer stays legal the whole way through. That sounds sensible, but it boxes the model in. It can't take a detour through bad territory to reach a better destination.

HardFlow drops that. It lets the intermediate steps wander, and steers the process so the final output lands inside the rules. They set it up as an optimization problem over the whole path, then broke it into one-step pieces to keep it cheap to run.

Two practical bits. It runs at deployment time on models you already have, so there's no retraining. And because it's an optimization, you can ask for more than legality, like a robot path that avoids collisions and is also short. They tested it on robotics, process control and computer vision.

Source: https://news.mit.edu/2026/new-method-enables-ai-safety-critical-situations-0914