How children acquire language in just a few years, from babble to grammar, remains one of the most compelling puzzles in cognitive science. Two challenges have long constrained progress: getting the data, that is, capturing language learning as it happens in children’s rich, messy, naturalistic environments, and making sense of it, via manual annotation that is slow, costly, and impractical at scale. Over the past decade, technological advances have begun to reshape both challenges at once. The transformation has been rapid and wide-ranging, spanning the hardware that captures children’s everyday environments, the algorithms that extract meaning from raw signals, and the scientific questions that can now be addressed as a consequence. In this review, we examine the latest work, published between 2024 and 2026, on ML both as a tool, enabling behavioral measurement at scales previously unattainable, and as a model, allowing researchers to instantiate and test explicit hypotheses about learning mechanisms.
Sense of agency is the subjective feeling of being in control over one’s own actions and their sensory outcomes. Here, we propose that a key feature of sense of agency is its link to intrinsic motivation, which is a key driver for engagement with the environment without external reinforcement. We argue that intrinsic motivation is highest when the sense of agency is experienced at an intermediate level between full control and no control at all. In consequence, we propose that human-operated, artificial intelligence-enabled robotics needs to be designed such that the balance between human sense of agency and the autonomy of the assistive system is achieved, for optimizing the users’ motivation to use the technology, and for improvement of performance. We describe various examples of assistive technologies, ranging from robots in work environments to healthcare robotics, and we discuss how sense of agency is a crucial factor that needs to be accounted for in the engineering endeavor of developing such technologies.
The hippocampus plays a critical role in generalization, enabling us to flexibly repurpose prior experiences to perform novel tasks. Here, we suggest that data augmentation — a machine learning strategy to improve generalization by refactoring prior experience — offers a useful framework to conceptualize and model hippocampal function. We begin by outlining how data augmentation operates across two timescales: the traditional ‘offline’ setting, where refactoring training data yields more general representations, and an ‘online’ setting, where retrieved experiences can be flexibly refactored at test time to support zero-shot inference. We suggest that these ‘offline’ and ‘online’ computational strategies map onto functions supported by the hippocampus. Critically, we argue that these computational tools can be leveraged to develop formal ‘linking functions’ between experimental evidence and theoretical claims, such that a unified modeling approach can be used to predict the diverse behaviors that depend on the hippocampus — from navigating in high-dimensional sensory environments to more abstract inferences. We hope that this perspective, and the modeling strategies it makes available, will support new efforts to formalize and evaluate theories of hippocampal function.
Planning is often described as the use of a world model to evaluate possible futures before choosing an action. In this review, we ask how such models are acquired, utilized, and revised in the brain. We frame the world model as the combination of a transition model, which predicts possible future latent states, and an outcome model, which assigns decision-relevant consequences to states and trajectories. We organize the review according to four complementary perspectives. Intuitive physics contributes to transition-model learning by providing a physics prior over object-centered latent dynamics. Model-free reinforcement learning contributes to outcome-model learning by providing structured teaching signals over reward distributions, task-relevant features or states, and temporal horizons. Successor representation supports flexible outcome-model updating by learning predictive maps that combine with reward functions to generalize value, adapt policy, and refine outcome estimates. Model-based reinforcement learning then deploys the acquired world model during deliberation and, through metacognitive control, evaluates whether it should continue to be queried, trusted, revised, or replaced. Together, these perspectives suggest that biological planning may depend not on a single monolithic representation, but on coordinated predictive resources that constrain possible transitions, shape outcome expectations, support flexible outcome update, and govern model usage. This framework also suggests that brain-inspired artificial intelligence should consider not only whether an agent has a world model, but what kind of predictive organization supports its use.
The rapid rise of large language models (LLMs) has sparked intense debate across multiple academic disciplines. While some argue that LLMs represent a significant step toward artificial general intelligence or even machine consciousness (inflationary claims), others dismiss them as mere trickster artifacts lacking genuine cognitive abilities (deflationary claims). We argue that both extremes may be shaped or exacerbated by common cognitive biases, including cognitive dissonance, wishful thinking, and the illusion of explanatory depth, which distort reality to our own advantage. By showcasing how these distortions may easily emerge in both scientific and public discourse, we advocate for a measured approach: a skeptical, open mind that recognizes the cognitive abilities of LLMs as worthy of scientific investigation while remaining conservative concerning exaggerated claims regarding their cognitive and moral status.
