What You Can't See Is What You Learn: Restricted Evidence Visibility Favors Compositional Generalization in Shared-Genome Language-Model Societies

TL;DR - Restricting each module’s access to evidence made shared-model, multi-cell language-model societies far more likely to learn compositional communication protocols than giving every module global visibility. The results suggest information bottlenecks can encourage reusable internal interfaces, though the preregistered evaluation narrowly failed its depth-three accuracy threshold.

• In 9 of 10 initialization-matched pairs, restricted-visibility societies outperformed globally visible counterparts by at least 20 points at both tested composition depths.
• Communication ablations reduced restricted societies to chance, while packet transplant and counterfactual experiments indicated reusable, value-indexed messages that causally controlled outputs.
• The depth-three advantage remained 0.558 for composite functions never seen during training, supporting genuine compositional generalization.
• Restricted-arm median depth-three accuracy was 0.6988, missing the preregistered 0.70 requirement; an earlier cohort also failed the complete battery due to poor ordinary-language preservation.