Constraining Generative Models: A Survey from the Constraint Programming Perspective

By Alexandre Bonlarron, François Pachet, Pierre Roy, and Jean-Charles Régin

Alexandre Bonlarron, François Pachet, Pierre Roy, and Jean-Charles Régin. “Constraining Generative Models: A Survey from the Constraint Programming Perspective”. Proceedings of IJCAI 2026, Survey Track, 7777–7786.

About the paper

A generative model can suggest a plausible next note or word. A constraint solver can track whether those choices satisfy requirements such as a specified length, a required pattern, or limits on copying. This survey looks at ways to combine the two.

Abstract in brief

The authors compare approaches that represent learned models as constraints, compile possible sequences into decision diagrams, or combine probabilities with constraint solving. The survey traces examples from Markov models and music composition to language models and discusses open questions in constrained sampling and inference.

Related research

The Markov Constraints project gives a musical example of this approach. The NeurIPS 2026 paper shows how to sample exactly under regular constraints for a variable-order Markov model.