Human recursive numeral systems (i.e., counting systems such as English base-10 numerals), like many other linguistic systems, seem to be shaped in similar ways, despite their inherent differences. Specifically, prior work from us (Prasertsom et al. 2025) have shown that, when compared to the space of plausible linguistic systems, which may or may not be attested, human systems tend to cluster together in a small part of the space of possible languages. This subspace is inherently comprised of those linguistic systems that are most optimal in terms of the trade-off between their regularity and processing complexity, informing us on what features might be more relevant for shaping the structure of these linguistic systems. This does not explain the motivation as to why these features might be relevant for these systems to arise instead of others. Following prior work that relates cross-linguistic tendencies to biases in learning, we are currently seeking if pressures in learning might bias these systems to be more regular and easier to process. We use reinforcement and supervised learning as a high-level analogue for human learning. We plan on doing additional work in several different directions: 1) How does use (as in, communicative use or use in terms of solving mathematical problems) complement learning in shaping these systems? 2) How might population sizes and contact between different populations might affect the evolution of these systems? 3) Can we generalise out of the domain of numeral systems to theorise what pressures on language might shape them, studying how systematic variations in the training dataset of large language models might influence the learnability of the systems they are tasked to learn? And could we explore the role of use in shaping the LLM's language too? 4) How does different paradigms of multilingual learning affect the proficiency and structure learned by LLMs? I will explore these questions along with my main supervisor Moa Johansson from Chalmers University and my second supervisor Devdatt Dubhashi, Chalmers University.