Program 2026

An influential theory in decision neuroscience posits a competition between a low-effort, “model-free” habit system and a complex, “model-based” goal-directed system. This talk presents behavioural and neuroimaging evidence that challenges this long-standing dual-system view, arguing that humans primarily use model-based inference when a task is properly understood.

We demonstrate that the widely-reported influence of model-free learning is likely an artefact of task misunderstanding. By providing clearer, story-based instructions for the classic two-stage task, we observe a dramatic shift to almost purely model-based choices. Crucially, this increase in correct model-based behaviour is accompanied by reduced mental effort and physiological markers of cognitive load, suggesting that standard abstract instructions promote a more cognitively costly, incorrect model-based approach rather than a less effortful model-free one. Furthermore, our neuroimaging analyses reveal that previous reports of combined model-free and model-based prediction errors in the ventral striatum are likely spurious, with more appropriate models showing no reliable evidence for a model-free signal.

These findings compel a re-evaluation of influential learning theories. We argue that what has been interpreted as low-effort, model-free behaviour may in fact reflect the higher cognitive cost of grappling with a misunderstood environment. This suggests the simple dichotomy between model-free and model-based systems is inadequate for explaining the complexities of human choice.

Every choice we make in life is accompanied by a sense of confidence – a subjective feeling about how likely it is that we are choosing the best possible course of action. Importantly, confidence is not only used to critically re-evaluate past decisions but also plays a pivotal role in guiding future behaviour when outcomes are not immediately available.

In this talk, I will present a psychophysical ‘dual-decision’ method that allows us to assess the calibration of confidence, revealing systematic biases of over- and under-confidence. I will then show how this approach can be used to separate confidence errors into bias and noise, uncovering domain-specific patterns: under-confidence in perceptual tasks, but increased confidence noise in knowledge-based tasks. These results shed light on the computational constraints shaping human confidence.

Decision-making is a fundamental cognitive process that underpins behavior across domains—from neuroscience and psychology to artificial intelligence and economics. Here, I will introduce the range of computational formalism and methods that are currently used in the psychology, neuroscience, and cognitive sciences. These will include reinforcement learning, Bayesian methods, machine learning techniques, and neural networks. I will show in particular how these models have been applied to explain specific cognitive phenomena and understand their neural correlates.

“Agentic AI” has recently become a popular term in discussions about the future of artificial intelligence, attracting interest from researchers, developers, and the public. But the idea of agents — systems that can sense, plan, and act in the world — is not new; it has been a central concept in AI, robotics, and cognitive science for decades.

This talk introduces the fundamental ideas behind agents, explains how modern “agentic AI” systems fit into this framework, and clarifies common misconceptions about what these systems can — and cannot — currently do. We will examine what it means for an AI system to be agentic, review examples of current agent architectures, and discuss major challenges such as aligning AI with human goals, making decisions under uncertainty, and ensuring long-term reliability. The goal is to equip students with a solid foundation for understanding and critically engaging with the rapidly evolving field of agentic AI.

Computational psychiatry is a maturing field that focuses on applying computational models of learning and decision-making to mental health. In this talk, I will highlight some findings from my group that underlie the diverse ways that modelling can be used to gain a mechanistic understanding of mental health concerns, including a meta-analysis of reinforcement learning approaches to depression and anxiety.

However, this approach is not without limitations – so I will go on to present data that provides a specific test of some of the implicit assumptions we make as a field, which has somewhat alarming implications – but also some promising ones.

Involving individuals with lived experience is increasingly being recognised as important in the study of psychopathology. In recognition of the value of collaborating with experts by experience, international funders are increasingly mandating that mental health research is developed by teams that include individuals from the population of study.

Yet, there is limited understanding of how to implement Patient and Public Involvement (PPI), including co-production and co-design, in fields such as computational psychiatry. Further, there is heterogeneity in how such methods are implemented and reported, limiting reproducibility. In this session, I will outline finings of a scoping review examining how PPI methods are implemented in mental health research, providing a practical introduction into how these methods can be adopted in computational psychiatry. I will also briefly outline emerging findings from a Delphi study into facilitators and barriers to diversity in PPI, which can help to ensure experts by experience who collaborate on research reflect the diversity of populations that experience mental health problems.

In decision-making, values attached to options can stem from past experiences with rewards and punishments (experiential) or explicit descriptions of outcomes and probabilities (descriptive). According to the common currency hypothesis, we encode these subjective values on the same scale. Most studies examine decision-making within either experiential or descriptive options separately, but what about when individuals choose between the two? Garcia et al. (2023) examined such hybrid choices by asking participants to choose between learned experiential options and symbolically described ones, and found systematic neglect of experiential options. This study explores, through five experiments, whether experiential neglect arises from differences in options’ and outcomes’ representations. Our findings show that experiential neglect persisted across all conditions, further challenging the dominant theory and suggesting that the neglect is primarily driven by memory retrieval cost. Furthermore, we found patterns of reduced sensitivity to losses, especially in contexts involving comparative decision-making.