Regularity Boosts Learnability of Recursive Numeral Systems

A new study using Reinforcement Learning methods confirms that regular recursive numeral systems, like English base-10 counting, are easier to learn than irregular but possible systems. The research, published as a preprint on arXiv (2602.21720), builds on prior work linking cross-linguistic tendencies to learning biases. The asymmetry favoring regularity emerges under the assumption that numeral systems are designed for generalization from limited data to represent all integers exactly. For highly irregular systems, learnability is influenced more by signal length than by regularity, suggesting different pressures apply across different parts of the system. The study provides computational evidence for why regularity is common in human languages.