🧠 neuroscience

Change-point oddball elicits pupil-linked signatures of Bayesian belief updating

This study introduces a "change-point oddball" protocol that demonstrates pupil-linked arousal signatures of Bayesian belief updating in response to unpredictable environmental changes, establishing a passive, no-report method for investigating adaptive learning across species and clinical populations.

Original authors: Kopel, L., Hoomans, J. M., van Gastel, E., Tijhuis, J. V., van de Graaf, M., Shafik, M., Matthijsse, S., Murphy, P. R., van Gaal, S., Suzuki, M., de Gee, J. W.

Published 2026-10-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kopel, L., Hoomans, J. M., van Gastel, E., Tijhuis, J. V., van de Graaf, M., Shafik, M., Matthijsse, S., Murphy, P. R., van Gaal, S., Suzuki, M., de Gee, J. W.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The human brain is a prediction machine. It constantly builds a model of the world based on what it has seen and heard before, using that model to guess what will happen next. When reality matches the guess, the brain stays calm. But when something unexpected happens—a sound that doesn't fit the pattern, a sight that breaks the rule—the brain must decide what to do. Is this just a random glitch, a momentary noise in the system? Or is it a sign that the rules of the world have actually changed? This split-second decision is the difference between ignoring a harmless mistake and updating your entire understanding of your surroundings. Getting this wrong can be costly, leading to confusion or anxiety, while getting it right is essential for navigating a shifting environment. Scientists have long used simple listening tests to study how the brain reacts to surprises, but those tests usually involve a stable world where the rules never change, making it impossible to see how the brain learns when the rules themselves are in flux.

A team of researchers at the University of Amsterdam and other institutions has developed a new way to watch the brain solve this problem without asking people to do anything but listen. They created a listening game where the rules flip unpredictably. Participants sat in a quiet room and listened to a stream of pure tones, some high and some low. For a while, one tone would be very common and the other rare. Then, without warning, the roles would swap: the rare tone became common, and the common one became rare. This happened at random moments, forcing the listener to constantly figure out which rule was currently in effect. The researchers did not ask the participants to press buttons or speak; they simply watched the participants' pupils. Even in a dimly room with constant light, the size of the pupil changes in response to the brain's internal calculations, acting as a window into the activity of systems that handle surprise and learning.

The study revealed that the pupil reacts differently depending on whether a sound is surprising because it is rare, or surprising because the world has changed. When a sound violated the current rule, the pupil dilated more than when a sound followed the rule. Crucially, this reaction was not just about the sound being different from the one before it. In a control experiment where ten different tones were played with equal frequency, the pupil did not show the same strong reaction to the "rare" sounds. This proved that the brain was not simply reacting to repetition or novelty, but was actively tracking a hidden state of the environment. The pupil response was largest when the brain was most uncertain about whether a change had occurred, and it faded away as the new rule became established and the brain grew confident again.

The researchers found that the speed at which the pupil reaction faded matched the speed at which a mathematical model of perfect learning would adjust its beliefs. This suggests that the physical expansion of the pupil is tightly linked to the brain's process of updating its internal map of the world. To confirm this link, the same group of participants later played a virtual reality game where they had to steer through gaps in a wall. The position of the gaps followed the same hidden rules as the tones. The study showed that people whose pupils reacted more strongly to the possibility of a rule change during the listening task were also the people who adjusted their steering behavior more quickly when the rules changed in the game.

This connection suggests that the same biological mechanism drives both the involuntary physical reaction of the eye and the conscious decision to change one's mind. The findings indicate that the brain uses a specific signal, likely involving a chemical messenger called noradrenaline, to mark moments when the world might have changed. This signal boosts the brain's ability to learn from new information, allowing it to discard old assumptions and adopt new ones. By showing that a simple, passive listening task can reveal these deep learning processes, the researchers have provided a new tool for studying how the brain adapts to change. This approach could eventually help scientists understand why some people struggle to update their beliefs in the face of new evidence, a difficulty seen in various mental health conditions, without requiring them to perform complex tasks or give verbal reports. The study demonstrates that even when we are just sitting still, our bodies are constantly working to figure out if the world has changed, and our pupils are the first to know.

Technical Summary: Change-point oddball elicits pupil-linked signatures of Bayesian belief updating

Problem Statement
The conventional auditory oddball protocol is a staple in cognitive and clinical neuroscience for studying neural responses to unexpected events. However, a fundamental limitation of this design is that stimulus statistics are stationary; deviants occur as rare events within a stable environment and do not signal a change in the generative structure of the environment. Consequently, the paradigm cannot distinguish whether a prediction error reflects transient noise or a genuine environmental change (a "change-point"). This distinction is critical for adaptive behavior, as the brain must decide whether to update its beliefs about latent environmental states or maintain current expectations. Existing active inference tasks (e.g., reversal learning) require overt responses, which confound belief updating with decision-making and motor preparation. There is a need for a passive protocol that preserves the simplicity of the oddball design while enabling the study of continuous, trial-by-trial Bayesian belief updating over latent states.

Methodology
The authors introduce "change-point oddball," a passive, no-report protocol designed to probe adaptive belief updating. The study involved three experiments with healthy human participants:

Experiment 1 (Passive Auditory/Visual Change-point Oddball): Twenty participants passively listened to sequences of pure tones (1 kHz and 2 kHz) or viewed Gabor patches. The stimuli were generated by one of two latent states (State A and State B) with reversed stimulus-probability mappings (e.g., in State A, 1 kHz is 90% probable; in State B, 2 kHz is 90% probable). State transitions occurred unpredictably with a hazard rate of 0.05. Pupil size was recorded at constant luminance to measure arousal-linked responses.
Experiment 2 (Many-Standards Control): Fifteen participants completed the change-point oddball protocol alongside a control condition featuring ten equiprobable tones. This design allowed the authors to dissociate prediction-related effects (oddball vs. control) from repetition-related effects (standard vs. control).
Experiment 3 (Cross-Task Association): Thirty participants completed an audio-visual version of the passive change-point oddball task and an active virtual reality (VR) change-point navigation task. In the navigation task, participants guided a character through gaps in walls, where the gap location was determined by latent states with a hazard rate of 0.08 or 0.12. The study aimed to correlate pupil-linked signatures of belief updating in the passive task with behavioral signatures of belief updating in the active task.

Computational Modeling
The authors employed a normative Bayesian belief updating model to derive trial-wise computational variables. Key variables included:

Posterior Belief: The probability of the current latent state given the history of observations.
Change-Point Probability (CPP): The posterior probability that a state change occurred on the current trial, derived from the subjective hazard rate and incoming evidence.
Uncertainty: Defined as the inverse of the absolute prior belief.

Key Results

State-Dependent Mismatch Responses: Pupil responses were significantly larger for state-specific oddballs compared to state-specific standards. This interaction confirmed that pupil-linked arousal tracks the current inferred latent state, not just physical stimulus properties.
Adjustment to State Changes: Following a state change, pupil responses to the newly established standard were initially large but decayed exponentially over subsequent presentations (time constant $\tau \approx 0.86$ standards). This decay mirrored the decay of model-derived change-point probability.
Prediction vs. Repetition: In the many-standards control experiment, the pupil mismatch response (oddball vs. standard) was primarily driven by prediction-related effects (84% of the total effect) rather than repetition suppression. This supports the view that the response reflects top-down prediction error.
Linear Tracking of Change-Point Probability: Pupil response amplitudes scaled linearly with model-derived change-point probability. This relationship held across both the dilation and constriction phases of the pupil response.
Cross-Task Association: Individual differences in the strength of change-point probability encoding in the pupil during the passive task significantly predicted the extent to which participants modulated their belief updates based on change-point probability during the active navigation task. Specifically, participants with stronger pupil encoding of CPP showed greater sensitivity to CPP when updating their positions in the VR task. This association was specific to the interaction between prediction error and change-point probability, not general prediction error or uncertainty.

Significance and Claims
The paper claims that the change-point oddball protocol successfully bridges the gap between the extensive oddball literature and contemporary work on probabilistic inference. By demonstrating that passive pupil responses encode signatures of Bayesian belief updating (specifically change-point probability), the authors argue that this protocol:

Preserves core signatures of adaptive belief updating without requiring overt behavioral reports.
Offers a scalable, non-invasive measure suitable for cross-species research (e.g., animal models) and clinical populations where extensive training or active participation is difficult.
Provides a physiological biomarker (pupil-linked arousal) that correlates with individual differences in normative belief updating strategies.
Supports the theoretical framework that phasic pupil-linked arousal, likely mediated by the locus coeruleus–noradrenaline system, plays a role in resetting computations and driving belief updates in volatile environments.

The authors conclude that while the protocol uses discrete stimuli (limiting the complexity of evidence integration compared to continuous distributions), it effectively isolates the mechanisms of deviance detection and latent-state inference, offering a powerful tool for investigating aberrant belief updating in conditions such as psychosis and anxiety.

More like this