Recent work suggests that language production exhibits a bias towards efficient information transmission. Speakers tend to provide more linguistic signal for meaning elements that are difficult to recover while reducing contextually inferrable (more frequent, probable, or expected) elements. This trade-off has been hypothesized to shape grammatical systems over generations, contributing to cross-linguistic patterns. We put this idea to an empirical test using miniature artificial language learning over variable input. Two experiments were conducted to demonstrate that the inferrability of plurality information inversely predicts the likelihood of overt plural marking, as would be expected if learners prefer communicatively efficient systems. The results were obtained even with input frequency counts of the plural marker counteracting the bias, and thus provide strong support for a critical role of inferrability of meaning in language learning, production, as well as in typologically attested variations.