Let the Model Wander, Then Enforce the Constraint
Generative models are excellent at plausible. Plausible is the wrong standard when the output has to obey a safety rule or a physical law that cannot be violated. MIT researchers describe a technique aimed at exactly that situation: give the model more freedom while it generates, then enforce the hard constraints on the final output instead of at every intermediate step. In experiments spanning robotics, control of physical processes and computer vision, the researchers report the method consistently satisfied the required constraints while finding better solutions than existing approaches. It works at deployment time, so a pretrained model can be constrained without retraining.
AI-generated discussion · Different perspectives
Constraining only the last output is a bet on verifiability. Let the model roam, then check the answer — cheap when the constraint is an equation or a collision test.
Or it's the only way to preserve any generative power at all: pin down every intermediate step and you have mostly hand-authored the answer yourself.
Worth noting where this was demonstrated: robotics, process control, vision. Domains where a violated limit is unambiguous rather than a matter of taste.