Inlay

Why do rituals persist? They are often treated either as attempts to influence uncertain outcomes or as practices that regulate group life. But these functions are usually blended together. In our new paper, we argue that reinforcement learning helps explain this blending.

Do LLMs *understand* language? Do educational AI agents *understand* the material they teach (or their students)? Claims about what AI systems do or don't understand are pervasive, but assessing them requires an account of MACHINE UNDERSTANDING

Bayesian efficient coding unifies two foundational theories of sensory processing: efficient coding and Bayesian inference. Central to this account is the idea that natural environmental statistics shape both how sensory information is encoded and how it is perceptually interpreted. By unifying these principles, the framework accounts for counterintuitive perceptual biases and establishes lawful relationships between environmental statistics, bias, and discrimination thresholds. In this article, we review behavioural and neural evidence for this theory in perception and cognition, as well as how short- and long-term adaptation to the environment may be expressed within the framework. We further review theoretical developments that extend the original framework, focusing on how response biases can be decomposed into encoding- and decoding-related components. A decade after its introduction, Bayesian efficient coding continues to evolve as a powerful theory, with recent extensions addressing early limitations and opening new directions for investigating perception and cognition.

From petitionary prayers to pilgrimages, rituals are found in every known culture. Yet, the reason for their persistence is a matter of active debate. Some studies portray rituals as attempts to affect uncertain outcomes, whereas others emphasize their role in facilitating social cohesion. We review the cognitive processes underlying both perspectives and draw on advances in reinforcement learning to integrate them. Specifically, ritual participation is motivated by two processes: habitual reinforcement of affective and social rewards experienced during performance (model-free learning) and reinforcement of pragmatic and cooperative benefits derived from culturally shared world models (model-based learning). This framework synthesizes previous accounts and illuminates ritual’s role in sustaining intersubjectively aligned world models in past and present societies.

Recent discoveries from Been et al., Schoenemann et al., and Verheijen et al. challenge our understanding of Neanderthals. Their brains appear more human-like than previously believed, and their hunting is cooperative and well organized. Yet their infants may have developed more rapidly. What made modern humans distinctive might be a prolonged childhood.

Recent work by Siems et al. shows that the brain rhythmically samples competing alternatives through covert spatial attention. This challenges continuous models of decision-making and suggests that evaluation is temporally structured by oscillatory dynamics, with attention determining when alternatives are accessed rather than reflecting changes in their representations.

Natural behavior unfolds as a continuous stream of actions. Because these actions often occur in rapid succession, the brain must prepare multiple future actions while the current action is being executed—a process we refer to as online planning. We review evidence for online planning in unpredictable movement sequences and consider neural implementations that could support parallel execution and multiple planning processes. We then show how this new framework could apply to the learning of specific sequences, improving performance while retaining the ability to modify sequences online. Online planning, therefore, provides a unifying account of how both unpredictable and well-learned sequences are produced, and how training leads to skillful and coordinated performance while retaining behavioral flexibility.

Human brain architecture is guiding brain-inspired artificial intelligence (AI) and has been treated as an optimal template, whose deviations could mark different psychiatric and neurological conditions. We argue this premise is wrong: under any single goal (e.g., minimal wiring cost or maximal communication efficiency), the human connectome is suboptimal. Instead, its organization reflects multi-objective trade-offs navigated over evolution and development under biological and environmental constraints. For psychopathology, atypical trajectories are not distances from an ideal brain but reweighted compromises in the same trade-off space. For neuro-AI, directly duplicating the brain’s connectivity risks copying its irrelevant compromises. Treating brains and models as products of multi-objective optimization and co-tuning relevant objectives offers a more powerful framework for interpreting clinical phenotypes and designing next-generation AI.

Attentional priority is typically conceived as a static spatial map, despite attention operating in a continuously changing world. We propose a dynamic priority map and outline the core demands and key questions needed to understand attentional guidance in a world that never stands still.

What do artificial intelligence (AI) systems “understand”? This question arises not only in assessing a system’s intelligence but also in evaluation practices to ensure the safe and responsible deployment of AI. Drawing on scholarship from philosophy and cognitive science, and informed by current practices in AI, we develop a framework for asking more precise questions and making more precise claims about machine understanding. We conceptualize understanding as a relation between a system (S) and a target of understanding (T), and we discuss how to specify the relation, the system, and the target, offering a landscape of options in each case. Our goal is not to defend a particular account of understanding, but to provide conceptual tools for those working to assess or advance machine understanding.