New research from the University of the Witwatersrand in South Africa shines a transformative light on the intertwined evolution of human language acquisition and artificial intelligence. By exploring the depths of “iterated learning,” a concept which suggests that language structure emerges and refines over successive generations, the study bridges linguistics, neuroscience, and machine learning in a novel and impactful way. This interdisciplinary breakthrough offers profound insights into how both children and AI systems develop increasingly coherent and systematic language understanding through gradual exposure and iteration.
The core premise of iterated learning builds on the idea that language is not static; it evolves dynamically through cultural transmission, where each generation learns from the previous one while introducing subtle modifications. Dr. Devon Jarvis, lead author and Lecturer at the School of Computer Science and Applied Mathematics (CSAM), explains that this research mimicked the processes of a child’s brain using deep linear neural networks—artificial systems designed to represent fundamental mechanisms of human cognition. Through this model, they observed that as generations of these ‘computer brains’ learned language data with intrinsic human-like properties, the structures within language naturally became more systematic and learnable over time.
Children acquire language and world knowledge in incremental, hierarchical stages, which frames the way they internalize concepts and categorize their environment. For example, they first differentiate between plants and animals and later understand more specific subcategories like different types of animals. This staged development explains fascinating anomalies in childhood cognition such as overgeneralizations: children may initially assume all birds can fly due to learning one characteristic but later adjust their understanding upon learning that penguins, though birds, do not fly but swim instead. According to Jarvis, such errors are not failures but integral steps in refining complex mental models through recursive learning and cultural transmission.
The research emphasizes that the imperfections in language transmission—mistakes, overgeneralizations, or simplifications—are critical in shaping language’s structure. These transmission errors act as a filter; simpler, more learnable pieces of language endure, while irregular or overly complex components tend to be lost or transformed. This selective pressure enhances communicative efficiency and consistency across generations and parallels the evolutionary forces that shape biological traits. Essentially, language is sculpted to be optimally learnable within the constraints of human cognitive architecture and social interaction dynamics.
A key revelation from this work lies in the nature of the neural networks used to simulate the learning process. The research team distinguished the differential impact of network depth—how many layers a neural system has—on the success of iterated learning. Their findings make it clear that deep linear networks, with multiple layers of hierarchical processing, effectively captured the emergence of compositional and systematic language structures. In contrast, shallow networks failed to replicate the nuanced, structured regularities that characterize human languages. This indicates that a crucial element for language acquisition, whether in humans or machines, is the ability to process information through layered, complex networks rather than flat or simplistic architectures.
This result has far-reaching implications for the development of artificial intelligence, particularly in the context of large-scale generative AI systems currently reshaping the technological landscape. Such models owe their emergent linguistic capabilities to the richness of their architecture and the complexity of their training environments. The research implies that the scalability in both network depth and training corpus complexity is not merely a matter of brute computational force but a foundational requirement for the natural evolution of structured and learnable language within AI.
Interestingly, the team connects long-established fields: deep linear networks have long been accepted as simplified but powerful models of child brain development, while iterated learning is a central linguistic theory explaining how language grows more ordered through cultural passage. The synergy of these perspectives unveils a fundamental principle—that language structure is not inherently given but is an adaptive product shaped by the sequential, hierarchical modes in which children learn and communicate. Learning stage by stage, children inherently favour reusing known patterns, which guides language to evolve towards higher compositionality and systematicity.
Moreover, the study’s success in using deep linear networks, a relatively simple form of neural network modeling, offers optimism in uncovering the fundamental cognitive principles underlying both natural and artificial intelligence. It suggests that despite AI’s increasing complexity, many of its core learning phenomena may be understood through elegant, interpretable mathematical constructs mirroring human cognitive development. This intersection of child development theories, linguistic iterated learning, and neural network architectures underscores the universal scaffolding guiding intelligent behavior across biological and synthetic realms.
The researchers note that while much work remains to generalize these findings beyond simplified models, their research presents a compelling argument for the crucial role of network depth and environmental complexity in language acquisition. This helps to transpose decades-old linguistic hypotheses into modern computational frameworks, equipping AI scientists and cognitive neuroscientists alike with new tools to probe the mechanics of learning and communication.
This exploration also revisits a classic debate in cognitive science regarding nature versus nurture in language development, showing that structured data environments and the constraints imposed by communication needs are indispensable factors complementing innate cognitive abilities. As AI systems advance, understanding how cultural transmission mechanisms and learner architecture jointly shape language evolution will be vital in designing programs that communicate with adaptive sophistication and nuanced understanding.
The collaboration behind this breakthrough brings together a distinguished team from several leading research institutions. Alongside Dr. Jarvis, co-authors include Professor Richard Klein, Head of the School of CSAM and a Fellow at the Wits MIND Institute, Professor Benjamin Rosman, Director of the Wits MIND Institute and researcher within CSAM, and Professor Andrew Saxe, affiliated with the Gatsby Unit and Sainsbury Wellcome Centre at University College London. Their combined expertise spans computational modeling, cognitive development, and linguistic theory, fostering a holistic approach to these complex questions.
Ultimately, this research illuminates how the seemingly disparate fields of child development, linguistics, and artificial intelligence converge to reveal underlying cognitive principles that govern learning, communication, and the emergence of language structure. It invites a reevaluation of how cultural and architectural factors shape intelligence and paves the way for novel insights not just into human cognition but the future trajectory of AI systems capable of genuine linguistic creativity and understanding.
Subject of Research: Iterated learning and language acquisition modeled through deep linear neural networks, exploring the co-evolution of language structure in humans and AI systems.
Article Title: Compositionality and Systematicity Emerge from Iterated Learning in Deep Linear Networks
News Publication Date: 5-May-2026
Web References: https://www.pnas.org/doi/10.1073/pnas.2509739123
References: Jarvis, D., Klein, R., Rosman, B., & Saxe, A. (2026). Compositionality and Systematicity Emerge from Iterated Learning in Deep Linear Networks. Proceedings of the National Academy of Sciences.
Keywords: iterated learning, deep linear networks, language acquisition, compositionality, systematicity, neural networks, AI language models, cognitive development, cultural transmission, linguistics, artificial intelligence, child development
