Evidence Masking

A design strategy for multi-module AI systems that restricts a module's access to evidence produced by other modules, thereby encouraging more compositional collaboration between modules and improving the system's overall compositional generalization capability.