AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-20
LogicVal:A Logic-Formalized Benchmark for Constraint Satisfaction in Text-to-Image Generation
To evaluate whether text-to-image models satisfy the constraints in a prompt, the authors recast evaluation as a constraint-satisfaction problem and formalized it in a logical framework. Across nine constraint classes and four models, the best model scored 98% on single-object constraints but only 53% on complex multi-object relationship constraints. They also propose LogicVal, which combines neural visual detection with logical reasoning and reaches ROC-AUC scores of 0.88 to 0.98 on CLEVR.