Predictive coding emerged from information theory, the mathematical framework Claude Shannon developed in 1948 showing that communication becomes efficient when redundancy is exploited. Video compression offers a familiar analogy: rather than transmitting every pixel of every frame, only the changes between frames are encoded. Predictive coding proposes that the brain does something similar. Instead of shuttling raw sensory data up the cortical hierarchy, the cortex transmits only the residual, unpredicted information—the prediction error—while higher levels send expectations downward. Perception, in this view, is a process of causal inference in which generative models of the world constantly compare their predictions against incoming sensations, minimizing the mismatch between internal model and external reality.

The architecture is elegantly hierarchical. At each level—say, orthographic, lexical, and semantic in the case of reading—two functionally distinct populations of neurons are hypothesized: state units encoding the current internal representation, and error units encoding information that deviates from top-down expectations. Predictions flow down the hierarchy through feedback connections, while errors flow upward through feedforward connections. Higher-level areas can effectively tell lower levels to “shut up” by explaining away sensory input, or to “stop gossiping” by sharpening responses inconsistent with the high-level interpretation. A precision-weighting mechanism modulates how much trust the system places in predictions versus bottom-up signals, allowing it to handle noisy or ambiguous input—much as Bayesian multisensory models explain the McGurk illusion, where an incongruent facial cue fuses with a phoneme into a hybrid percept.

To evaluate whether this framework applies to language, the authors adopted four hypotheses previously used to assess predictive coding in visual perception: that error-signaling neural responses should scale inversely with expectation; that top-down signals represent sensory prediction; that distinct neural subpopulations encode predictions and errors at each cortical level; and that prediction error minimization proceeds through reciprocal exchange of signals across hierarchical levels. The verdict, drawn from electrophysiology, eye-tracking, and behavioral studies, is a nuanced mix of confirmation and challenge.

The strongest support concerns the N400, the brain’s signature electrical response to meaning, a negativity peaking roughly 400 milliseconds after a word appears. N400 amplitude reliably decreases as a word’s cloze probability—how predictable it is from context—increases, a pattern first documented by Marta Kutas and Steven Hillyard in 1980. In predictive coding terms, expected inputs generate smaller prediction errors, which translate into attenuated N400 amplitudes. Recent computational work by Samer Nour Eddine and colleagues successfully simulated a wide range of N400 effects, including graded context effects for both orthographic and semantic manipulations, within an explicit predictive coding model.

Perhaps most striking is how the framework handles “related anomaly” effects. In a classic paradigm, readers see sentences where an unexpected word is either from the predicted category or not: after a setup about playing Monopoly, the word “chess” elicits a smaller N400 than “baseball,” even though both are equally implausible. Predictive coding models can explain this through top-down lexico-semantic facilitation—semantic overlap between predictions and sensory input suppresses both lexical and semantic prediction errors. Similarly, parafoveal N400 effects, where the brain responds to words viewed in the periphery before direct fixation, scale with contextual expectation, and these effects attenuate once the word is fixated, consistent with prediction error already having been partially minimized.

Behavioral evidence from eye-tracking also fits the framework reasonably well. Predictable words are skipped more often, fixated more briefly, and yield larger preview benefits—findings replicated since the early 1980s. Predictive coding models can accommodate these effects if behavioral response times reflect the settling of internal state representations rather than prediction error per se, an interpretation that parallels evidence accumulation models like the drift diffusion framework. The architecture’s implicit multi-scale processing also offers a way to understand why predictability benefits appear in both first-pass and later eye movement measures, something single-threshold models of eye movement control struggle to explain.

But the picture darkens at the edges. The frontal positivity, an ERP component elicited roughly 500 to 1,000 milliseconds after plausible prediction violations—unexpected words in highly constraining contexts—proves stubbornly difficult to reconcile with predictive coding. Unlike the N400, this component does not reliably show graded responses to unexpected input, and its elicitation conditions remain inconsistent across studies. Some work finds it only when predictions are actively violated; other work finds it regardless of sentential constraint. The authors suggest it might reflect top-down bias generated when situation models are updated, but explaining why this activity is unique to plausible rather than implausible violations remains an open problem.

Biological plausibility presents its own challenges. Recent neurophysiological work in mice and monkeys found no evidence for the feedforward prediction error signals that predictive coding requires even in visual cortex, raising questions about whether such circuitry could support the distributed and heterogeneous language network. Meanwhile, oscillatory evidence offers partial support: anticipatory alpha suppression before expected words in constraining contexts is consistent with precision weighting, and studies of speech processing show hierarchical temporal organization across frequency bands. But the field lacks studies explicitly designed to dissociate prediction from prediction error in language, and some findings suggest cortical tracking of speech may be dominated by transient evoked responses rather than the narrow-band oscillatory entrainment popular in predictive coding accounts.

The authors conclude that predictive coding has genuine explanatory utility for linguistic processing—particularly for N400 context effects, anticipatory semantic overlap, and parafoveal-to-foveal attenuation—but likely does not constitute the sole core mechanism of the language architecture. They caution that predictive coding is inherently difficult to falsify: the precision-weighting mechanism can simultaneously elicit or suppress neural activity depending on context, making it hard to determine whether any individual effect supports top-down or bottom-up processing. Evidence for prediction, they note, is not automatically evidence for predictive coding, which is a specific model of neural efficiency rather than a broad framework. The most productive path forward may be a more accommodative view in which predictive coding slots into existing computational models of reading as one mechanism among several, rather than serving as the unified theory of everything the brain does with language.

Subject of Research: Testing predictive coding as a neural model of linguistic predictability effects in reading

Article Title: Evaluating the compatibility of predictive coding as a model of linguistic predictability effects in reading

Article References: Copeland, A., & Payne, B. R. (2026). Evaluating the compatibility of predictive coding as a model of linguistic predictability effects in reading. Frontiers in Cognition, 5, Article 1768534. https://doi.org/10.3389/fcogn.2026.1768534

Image Credits: AI Generated

DOI: 10.3389/fcogn.2026.1768534

Keywords: predictive coding, reading, N400, prediction error, psycholinguistics, eye movements, ERP, language comprehension, hierarchical inference, precision weighting, frontal positivity, computational modeling

The core problem Kafkas identifies is one that has frustrated cognitive neuroscientists for years. Animal studies have delivered exquisitely detailed accounts of how dopamine and norepinephrine shape hippocampal plasticity, influencing long-term potentiation, memory persistence and the prioritisation of behaviourally significant events. Work on reward, novelty and surprise suggests that catecholaminergic modulation is not one unitary arousal response but a family of state-dependent mechanisms through which the brain selects experiences for long-term storage. Yet translating those mechanistic insights into human research remains strikingly difficult, because direct measures of endogenous catecholaminergic activity in a living human brain are nearly impossible to obtain during the fleeting seconds in which memories are actually formed.

Existing tools each capture only a partial slice of the underlying biology. Positron emission tomography offers receptor- and transporter-specific measurements, but its temporal resolution is poor, it is costly and it involves radiation exposure. Neuromelanin-sensitive magnetic resonance imaging reveals the structural integrity of catecholaminergic nuclei, most famously the locus coeruleus, the brain’s principal norepinephrine factory, but says nothing about their moment-to-moment functional dynamics. Pharmacological manipulations can establish causal links between neurotransmitters and cognition, but they lack anatomical specificity and are confounded by systemic effects scattered throughout the body. What the field needs, Kafkas argues, is a biomarker that is non-invasive, inexpensive, temporally sensitive, easy to combine with imaging and modelling, and capable of tracking trial-by-trial fluctuations in neuromodulatory state.

Pupillometry ticks every one of those boxes, and an expanding literature shows that pupil dynamics genuinely matter for memory. Encoding-related pupil responses predict later recognition, memory strength, associative learning and mnemonic discrimination across a wide range of paradigms. They even distinguish qualitative features of remembering, separating recollection from familiarity, novelty from known, and in one striking study true memories from false ones. These are not trivial byproducts of effort or alertness, Kafkas contends, but physiological signatures associated with whether an experience gets encoded at all. Crucially, pupil responses also track the computational variables at the heart of hippocampal learning: novelty, expectation and prediction error all modulate pupil behaviour during encoding, and those modulations predict different subsequent memory outcomes.

The provocative part of the new framework is its refusal to treat the pupil as a narrow proxy for locus coeruleus activity. Pupillary responses emerge through the coordinated action of sympathetic and parasympathetic pathways and are shaped by distributed cortical, subcortical and brainstem circuits. Recent simultaneous pupillometry-fMRI evidence supports that broader view. During incidental encoding, pupil responses covaried with hippocampal and parahippocampal activity and predicted later recognition, and successful encoding was associated with pupil-related functional interactions between the hippocampus, the Edinger-Westphal nucleus, the ventral striatum and the orbitofrontal cortex. Pupil-linked memory effects, in other words, ride on a network spanning hippocampus, cortex and brainstem, not on a single neuromodulatory nucleus. Kafkas is careful about causality: these associations do not establish the direction of influence, and they may reflect direct interactions or common upstream drives, or both.

The framework’s boldest proposal tackles an apparent contradiction in the literature. Many studies report that larger pupil dilations during encoding predict successful memory. Others have found that encoding-related pupil constriction can predict memory too, under specific conditions. Rather than dismissing one set of findings, Kafkas suggests that constriction and dilation may index genuinely different physiological states recruited by different encoding demands. In an expectation-learning experiment, unexpected novel stimuli produced steadily increasing dilation and preferentially boosted later recollection, while expected novel stimuli elicited constriction yet still predicted successful recognition. Both routes led to memory, but through different physiological configurations.

Those two configurations map neatly onto two computational modes of hippocampal learning. When new information broadly fits existing knowledge, successful encoding may depend on integrating it into established representations, a state that plausibly leans more on parasympathetic and autonomic brainstem processes, consistent with the link between expected novelty, constriction and hippocampal connectivity with the Edinger-Westphal nucleus, the parasympathetic hub controlling the pupil’s constrictor muscle. When events violate predictions and demand revision of internal models, encoding may instead rely more heavily on catecholaminergic hippocampal-midbrain interactions. Here dilation during surprise reflects rapid memory updating rather than mere arousal, an idea compatible with the GANE model, in which norepinephrine amplifies high-priority neural representations under salient conditions, but extending it to hippocampal-midbrain network states. Animal work adds a further nuance: rapid pupil fluctuations track noradrenergic activity closely, whereas sustained dilation during locomotion tracks cholinergic activity instead. One pupil response, multiple possible neuromodulatory backdrops.

Turning the framework into hard science will require multimodal and, ultimately, causal tests, and Kafkas lays out a concrete experimental agenda. High-resolution fMRI can relate trial-by-trial pupil fluctuations to hippocampal subfields and network connectivity. Neuromelanin-sensitive MRI can index individual differences in catecholaminergic nuclei, despite ongoing challenges in accurately localising the locus coeruleus. Pharmacological studies can test whether cholinergic, dopaminergic and noradrenergic manipulations differentially alter constriction- and dilation-linked memory states. Magnetic resonance spectroscopy can probe the excitatory-inhibitory balance of the hippocampus, and autonomic measures such as heart-rate variability and skin conductance can help tease apart sympathetic and parasympathetic contributions. Computational modelling supplies the final ingredient: instead of simple remembered-versus-forgotten contrasts, future studies should estimate latent variables like surprise, uncertainty and belief updating on each trial, then relate them simultaneously to pupils, hippocampal activity and later memory quality.

The falsifiable predictions are explicit. Manipulating expectation while holding stimulus novelty constant should shift encoding-related pupil responses between constriction-dominated and dilation-dominated profiles, and those shifts should be accompanied by changes in hippocampal-midbrain connectivity rather than locus coeruleus activation alone. Drug studies should perturb the two profiles differently if they truly reflect distinct neuromodulatory states. The framework would falter, by contrast, if a single arousal process could explain everything, or if the dissociations failed to replicate beyond the small set of paradigms that currently motivate them. Kafkas acknowledges this limitation candidly, calling for independent replication and cross-species work in which circuit-level manipulations could determine whether pupil profiles are causally tied to specific neuromodulatory mechanisms or merely parallel consequences of broader physiological states.

The stakes extend well beyond the laboratory. Altered catecholaminergic signalling and hippocampal dysfunction are central features of healthy ageing, Alzheimer’s disease and several neuropsychiatric conditions, and recent evidence links pupil dynamics to functional network integrity in ageing and to Alzheimer’s disease biomarkers. Because pupillometry is cheap, non-invasive and endlessly repeatable, it could help characterise individual differences in neuromodulatory function during memory formation, not as a standalone diagnostic but as one layer in multimodal profiles combining memory performance, network integrity and disease progression. Whether every specific hypothesis survives testing or not, the reframing itself may prove durable: by treating the pupil as a systems-level marker of integrated neuromodulatory, autonomic and hippocampal processes, the field gains a scalable route for connecting molecular mechanisms of dopamine and norepinephrine to the human memories that depend on them, across health, ageing and disease.

Subject of Research: Pupillometry as a systems-level biomarker of neuromodulatory influences on human hippocampal memory

Article Title: Beyond the locus coeruleus: pupil dynamics as a systems-level marker of neuromodulatory influences on hippocampal memory

Article References: Kafkas, A. (2026). Beyond the locus coeruleus: pupil dynamics as a systems-level marker of neuromodulatory influences on hippocampal memory. Frontiers in Molecular Neuroscience, 19, Article 1976085. https://doi.org/10.3389/fnmol.2026.1976085

Image Credits: AI Generated

DOI: 10.3389/fnmol.2026.1976085

Keywords: pupillometry, hippocampus, dopamine, norepinephrine, locus coeruleus, episodic memory, neuromodulation, novelty, prediction error, memory encoding, autonomic nervous system, fMRI

The study, published in Frontiers in Neurorobotics by Pierre Klintefors, Birger Johansson, and Christian Balkenius, draws its inspiration from predictive processing, a leading theory of how the brain works. In that framework, perception and movement are unified under a single goal: minimizing the gap between predicted and observed sensory input. Movement itself is framed as a set of proprioceptive predictions—predictions about what the body will feel—rather than explicit motor commands. Sensory signals that match those predictions can be attenuated or ignored; signals that diverge demand attention. The Lund team asked a deceptively simple question: instead of treating prediction error as waste to be minimized, can a robot mine it for information?

Their answer came from Epi, an in-house two-degree-of-freedom robotic head built for developmental robotics research. The head is actuated by two Dynamixel MX-106R servomotors, one controlling pan and one controlling tilt, and carries an onboard UM7 inertial measurement unit with accelerometers and gyroscopes. Crucially, the team deliberately withheld tactile and visual feedback. The robot’s only windows onto the world were its own joint positions, its inertial readings, and the electrical current drawn by its servos—a value that rises and falls with joint load, making it a natural proxy for force. This constrained sensorium mirrors, in a rough way, the proprioceptive signals that human muscle spindles deliver, and it allowed the researchers to isolate just how much can be inferred from the body’s internal senses alone.

The first step was to give the robot an internal generative model—a learned sense of its own physics. The researchers collected training data through a process directly analogous to infant motor babbling: 2,000 randomized servo transitions, each separated by a 500-millisecond pause, with proprioceptive and inertial data sampled at 40 hertz throughout. From these recordings they trained windowed multilayer perceptrons to predict the servo current at the next time step from histories of position, distance-to-goal, inertial signals, and past current values. Each network stacked four fully connected hidden layers of 256, 128, 64, and 32 units, with LeakyReLU activations, batch normalization, L2 regularization, and dropout. Training used the Adam optimizer with a Huber loss chosen for robustness against outliers, and every configuration was trained three times with different random seeds to guard against initialization luck—54 training runs in all.

The architecture choice was itself a pragmatic engineering decision. While recurrent networks like LSTMs handle sequences elegantly, their step-by-step execution can bottleneck real-time deployment on modest hardware. A windowed MLP evaluates its entire temporal window in a single feedforward pass, which made it easy to convert to TensorFlow Lite and run on the robot’s MacBook-based control system at a strict 40-hertz inference rate. When the team compared their networks against linear ARX baselines given identical lagged inputs, every MLP variant won at every history length tested—a clear sign the networks were exploiting nonlinear structure in the sensorimotor stream rather than simply extrapolating a linear trend. One striking finding: giving the network just one step of current history (25 milliseconds) was enough, and performance stayed flat out to fifteen lags. Even more tellingly, once the model had three or more time steps of proprioceptive history, the inertial measurement unit became redundant. With enough short-term memory, the network could implicitly derive velocity and acceleration on its own—no specialized vestibular sensor required.

With the best model deployed in real time on Epi, the perturbation experiments began. The team tested five conditions: a brief manual push applied to the stationary head, a sustained force held throughout a trial, a rigid obstacle that intercepted the head during a 30-degree movement, and two unperturbed baselines—a gentle 5-degree movement and a full 30-degree sweep. Fifty trials per condition per servo yielded 500 classified trials. The manual pushes were deliberately varied in amplitude, ensuring the categories represented a range of real-world forces rather than one calibrated jab. From each two-second trial, the researchers distilled the resulting prediction error signal into twenty features: magnitude statistics such as maximum absolute error and cumulative error, temporal descriptors such as time-to-peak, half-decay time, lag-1 autocorrelation, and a temporal center of mass, and morphological shape coefficients drawn from the discrete cosine transform of the normalized signal.

The resulting profiles were vividly distinct. A brief push produced a sharp spike in prediction error that stabilized within about 400 milliseconds. Sustained force left a long, persistent tail in the error signal. An obstacle collision drove the error amplitude far above that of an identical unobstructed movement. The two unperturbed movements resembled each other, with larger amplitude peaks for the wider 30-degree sweep. A random forest classifier, trained under rigorous nested cross-validation with tilt and pan data kept in separate trials to prevent information leakage, achieved 98.5 percent overall accuracy across the five conditions, with a 95 percent bootstrap confidence interval spanning 97.4 to 99.6 percent. A permutation test confirmed the result was no statistical fluke: across 100 label-shuffled repetitions, none came close to the observed accuracy, yielding an empirical p-value of 0.0099.

Perhaps the most philosophically interesting result concerned what the classifier actually needed. Magnitude features alone reached 95.9 percent accuracy. But the amplitude-normalized temporal and shape features—those carrying pure temporal structure, stripped of how big the errors were—performed even better at 97.6 percent. The two most discriminative features were the temporal center of mass of the error signal and the mean error in the first of five phase bins, together accounting for over a quarter of the classifier’s importance weight. And when the researchers classified the raw servo current directly, bypassing the prediction error entirely, they got exactly the same 98.5 percent. In other words, the prediction error carried just as much diagnostic information as the raw signal it was derived from. The mismatch between expectation and reality turned out to be as informative as reality itself.

Why does this matter beyond a lab in Lund? The team suggests that error profiles could directly drive behavior in future systems: a sharp spike characteristic of a push might trigger reorientation or stabilization, sustained force might trigger compliance so the robot yields gracefully, obstacle contact might trigger an immediate halt, while expected movement-related errors would be attenuated and ignored. This resonates with sensory attenuation in biology, where self-generated sensations evoke weaker neural responses than externally caused ones—why you cannot tickle yourself, in essence. It also echoes classical control theory: the researchers blended predicted and actual current at a fixed 80-to-20 ratio, a design choice analogous to the uncertainty-weighted gain of a Kalman filter, keeping the error signal stable without drift.

The authors are candid that this is a proof of principle in a deliberately constrained setting. Perturbations came from a single operator pushing in a consistent direction, obstacle interactions occurred only during fixed 30-degree movements, and the platform has just two degrees of freedom. Scaling to robots with many coupled joints—where one limb’s load depends on the state of others—remains an open challenge, as does handling perturbation directions, magnitudes, and movement contexts beyond those tested. Hand-crafted features may also need to give way to learned temporal embeddings for messier interactions. Yet the conceptual payoff stands: imperfect prediction need not be a liability. The very structure of a robot’s surprise—its onset, its peak, its decay, its persistence—encodes what the world just did to it. In the predictive processing view the authors embrace, error is not merely a signal to silence but a resource to read, and machines that learn to read it may one day sense their world through their bodies alone, much as we do.

Subject of Research: Temporal profiles of load-related prediction errors for classifying physical perturbations on an embodied robot

Article Title: Temporal signatures of prediction error reveal external forces on an embodied robot

Article References: Klintefors, P., Johansson, B., & Balkenius, C. (2026). Temporal signatures of prediction error reveal external forces on an embodied robot. Frontiers in Neurorobotics, 20, Article 1918275. https://doi.org/10.3389/fnbot.2026.1918275

Image Credits: AI Generated

DOI: 10.3389/fnbot.2026.1918275

Keywords: cognitive robotics, predictive processing, prediction error, proprioception, perturbation classification, embodied robotics, servo current, motor babbling, active inference, neurorobotics, sensory attenuation, machine learning

The research team, led by neurosurgeon and neuroscientist Eyiyemisi Damisah of Yale University, took advantage of a rare clinical opportunity. Seven patients with medication-refractory epilepsy were undergoing robotic implantation of depth electrodes to localize the origins of their seizures. Each of these hybrid electrodes carried bundles of eight microwires, each just 40 micrometers across, protruding four millimeters from the electrode tip. Because the electrodes were placed according to clinical need, some of them happened to pass through the claustrum, the dorsal anterior cingulate cortex, and the basolateral amygdala, giving the researchers an extraordinary window into three brain regions at the resolution of individual neurons in awake, behaving humans.

While the electrodes recorded the firing of single neurons at 30,000 samples per second, the patients played a deceptively simple computer game. They steered a spaceship along a vertical axis, trying to slip through a safety hole in an approaching belt of asteroids. Two zones, one above and one below, each carried a probability of containing a safe passage that shifted unpredictably between blocks of trials, sometimes offering a 90 percent chance of safety and sometimes only 10 percent. Because the asteroids approached too quickly for reactive dodging, participants had to anticipate where the hole would appear based on what they had learned, positioning their ship in advance according to their internal beliefs about the hidden probabilities.

This task design was the key to the study’s ambition. The researchers fitted each participant’s behavior with an approximate Bayesian model, a mathematical framework in which beliefs about the safety of each zone are represented as probability distributions that are updated trial by trial. From this model, the team could extract two so-called latent variables, quantities that exist in the mind but are never directly observable: subjective uncertainty, derived from the variance of the belief distribution, and prediction error, the mismatch between what the participant expected and what actually happened. The model outperformed standard reinforcement-learning alternatives, and crucially, the uncertainty estimates had a visible behavioral signature. On trials when the model inferred high uncertainty, participants moved their spaceship more and with greater variability, as though exploring between the two potential safety zones.

When the researchers examined the neural recordings, the claustrum proved to be intensely engaged. Of the 110 single neurons recorded there, 78, or 71 percent, showed significant task-related firing changes, a proportion matched almost exactly by the anterior cingulate cortex, where 56 of 80 units were task-responsive. The amygdala, by contrast, was far less involved, with only about 30 percent of its neurons modulated by the task. Within the claustrum, distinct subpopulations emerged. Some neurons fired sharply and transiently at the moment the asteroids appeared, splitting into cells that burst into activity and cells that briefly paused. Others responded only to the outcome of each trial, and strikingly, these outcome neurons showed a strong bias toward crashes: they fired far more vigorously when the ship was destroyed than when it successfully avoided the asteroids.

That crash bias set the claustrum apart from its cortical partner. In the anterior cingulate cortex, outcome-responsive neurons were almost evenly split between those preferring crashes and those preferring successful avoidance, producing no consistent directional preference across the population. A linear mixed-effects model confirmed a statistically robust interaction between brain region and outcome type, and the effect survived leave-one-subject-out refits and log transformation of the firing rates. In other words, the two regions, though similarly engaged, appeared to be doing meaningfully different things with the same stream of task events.

The most striking findings emerged when the researchers linked neural activity to the model-derived latent variables. Roughly 28 percent of claustral neurons were modulated by uncertainty, firing more strongly on high-uncertainty trials during the active avoidance period, and about 23 percent responded to high prediction errors, with the two populations overlapping substantially. Because uncertainty and prediction error naturally covary, errors tend to be larger when beliefs are imprecise, the authors are careful to note that the analyses cannot fully separate the two signals at the single-neuron level. But the pattern is coherent: claustral neurons preferentially ramp up their activity when the brain’s internal model is least confident and when outcomes violate expectations most sharply.

To rule out the possibility that these signals were merely echoes of visible task events, the team applied a conditional mutual information analysis, a technique from information theory that measures how much a neuron’s firing depends on a latent variable after statistically controlling for confounds such as the ship’s position, the safety zone, and the outcome. During the active avoidance period, a significant fraction of claustral neurons retained information about uncertainty even after these controls, something that could not be explained by the external stimuli alone. The anterior cingulate cortex showed a complementary temporal profile: its uncertainty-related signals were most prominent during the intertrial interval, before the asteroids appeared, consistent with a role in maintaining and monitoring the reliability of internal beliefs between episodes of action.

The authors propose an elegant synthesis within the framework of predictive processing, the influential theory that the brain constantly generates predictions about the world and updates them based on errors. In this account, the anterior cingulate cortex may carry precision signals, estimates of how much the brain should trust its current beliefs, that tonically suppress claustral output during stable periods. When uncertainty rises, that suppression weakens, releasing bottom-up error signals from the claustrum to propagate through its vast cortical network and drive rapid updating. This would explain why claustral crash responses were strongest precisely under high uncertainty, and it fits with rodent work showing that the claustrum decides whether prediction-input mismatches are dismissed as noise or broadcast as meaningful news.

The implications reach well beyond basic neuroscience. Structural and functional abnormalities of the claustrum have been reported in conditions marked by disturbed perception and inference, including autism, schizophrenia, and psychedelic states, and the new findings suggest a concrete computational role that such abnormalities could disrupt. The study has limitations the authors acknowledge candidly: the cohort was small, the participants had epilepsy, and the number of neurons, particularly in the amygdala, was limited. Still, the demonstration that human claustral neurons track hidden cognitive variables, quantities no sensory input could directly convey, transforms this long-mysterious structure from a speculative curiosity into a serious candidate for a key node in the brain’s inferential machinery. Future experiments, including targeted microstimulation of claustral populations, may soon reveal whether this thin sheet of neurons truly acts as the conductor the theory predicts.

Subject of Research: Single-neuron encoding of uncertainty and prediction error in the human claustrum during aversive learning

Article Title: Human claustrum neurons encode uncertainty and prediction errors during aversive learning

Article References: Hu, M., Dalvit, R., Medina-Pizarro, M., Dougherty, M., Zhou, Y., Barreto-Nieves, J., Obaid, S., Afrasiyabi, A., Krishnaswamy, S., Kaye, A. P., Günel, M., Krystal, J. H., Sheth, K. N., Gu, X., Pittenger, C. P., Wise, T., & Damisah, E. C. (2026). Human claustrum neurons encode uncertainty and prediction errors during aversive learning. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02475-x

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02475-x

Keywords: claustrum, uncertainty, prediction error, aversive learning, anterior cingulate cortex, amygdala, Bayesian modeling, single-unit recordings, predictive processing, epilepsy patients, neuroscience, cognitive control

The study, led by Xinru Huang, Yinmei Ni, Yuxi Wang, Yujia Peng and Jian Li, drew on a framework that has quietly revolutionized computational psychiatry. Reinforcement learning describes how agents — animals, humans, or machines — improve their choices by tracking the difference between expected and received outcomes, a signal known as prediction error. The efficiency of that updating is captured by a parameter called the learning rate: high learning rates mean that each new outcome rapidly rewrites expectations, while low learning rates mean the learner integrates information slowly, clinging to older estimates. By fitting this model to behavior in a simple gambling-style task, scientists can read out cognitive parameters that no questionnaire can directly measure.

Participants in the study performed a probabilistic instrumental learning task, repeatedly choosing between options to either earn rewards or avoid losses. Two independent experiments were conducted, the first with 190 participants and the second with 361, giving the team a substantial sample on which to test whether personality-level differences in mood translated into measurable differences in learning. Alongside the task, participants completed classic psychiatric instruments: the Self-rating Depression Scale and the State-Trait Anxiety Inventory, Trait version. These questionnaires provided continuous measures of depressive and anxious traits across the full population, rather than forcing people into diagnostic categories.

The central result was a clean dissociation. Depression traits correlated negatively with learning rates: the more depressed a participant reported feeling, the more slowly their beliefs updated in light of new evidence. Anxiety traits showed precisely the opposite pattern, with higher trait anxiety predicting faster updating. Crucially, this mirror-image relationship appeared in both experiments, in tasks involving both gaining rewards and avoiding losses. The finding suggests that the two disorders, whatever they share at the level of mood, push the brain’s inference machinery in opposite directions — one toward sluggish conservatism in the face of new outcomes, the other toward rapid, possibly over-reactive revision of expectations.

But the researchers went a step further, asking which specific symptoms within the broad constellations of depression and anxiety were actually responsible for the effect. Depression is not a single thing: it bundles somatic complaints, cognitive difficulties and anhedonia — the loss of pleasure and interest — under one label. Using a transdiagnostic latent factor approach, the team decomposed the questionnaire data into these finer-grained dimensions and related each to learning behavior. The analysis revealed that somatic symptoms and anhedonia were the main drivers of the negative association between depression and learning rate. In other words, it is the bodily sluggishness and the inability to feel pleasure — not depression as an abstract whole — that accompany a slowed grip on new information.

Equally revealing was what emerged on the anxiety side. Cognitive symptoms and negative affect, the dimensions most central to anxious experience, correlated positively with learning rates. Anxious individuals, it appears, do not simply react more to the world; they absorb it faster, folding each surprising outcome into their model of the environment almost immediately. This makes intuitive sense in light of long-standing theories portraying anxiety as a state of heightened vigilance and threat anticipation, but the new results give that portrait a precise quantitative form: faster belief updating in the face of prediction errors.

The authors propose that the dissociation may reflect a fundamental trade-off between internal and external focus. Depression, particularly through anhedonia and somatic burden, draws attention inward — toward bodily states, rumination and self-referential thought — leaving less capacity for processing what the external environment is signaling. Anxiety, by contrast, is oriented outward and forward, scanning for danger and updating rapidly on anything that might matter. Under this account, the opposing learning rates are not incidental quirks but visible consequences of where each condition places the mind’s limited resources. Excessive internal focus, in this framing, directly diminishes the capacity for external information processing.

The work carries implications well beyond the laboratory. Diagnosis in psychiatry remains largely based on self-reported symptoms, yet two patients with identical scores can differ enormously in how they think and make decisions. Computational measures such as learning rates offer a way to characterize the underlying machinery rather than the surface presentation. If depressive learning slowdown is driven specifically by anhedonia and somatic symptoms, then those symptoms might serve as markers for a particular cognitive profile — and perhaps as targets for interventions designed to re-engage patients with the reward structure of their environment. Conversely, the accelerated updating seen in anxiety could illuminate why anxious individuals so readily revise threat expectations and, in some cases, develop persistent worry.

The transdiagnostic strategy at the heart of the study also answers a growing call in psychiatry to look past diagnostic categories. Rather than comparing a depression group against an anxiety group — an approach complicated by heavy comorbidity — the researchers treated mood traits as continuous dimensions present to varying degrees in everyone, then used latent factor modeling to isolate which symptom dimensions carried the behavioral signal. The result is a picture in which shared and distinct mechanisms can be quantified within a single unified computational framework, precisely the kind of rigor that proponents of computational psychiatry have long promised.

There are, of course, limits to what a laboratory task can capture, and the authors note that their findings reflect trait-level variation in generally healthy samples rather than clinical diagnoses. Future work will need to test whether the same dissociation holds in diagnosed populations, whether it predicts treatment response, and how the behavioral parameters map onto neural circuitry. Still, the study demonstrates something rare in mood research: a simple, mathematically defined parameter that moves in opposite directions for two of the world’s most common mental health conditions. In the quiet arithmetic of prediction errors and belief updates, depression and anxiety — so often mistaken for each other — finally tell two different stories.

Subject of Research: Computational dissociation of depression and anxiety through reinforcement learning mechanisms

Article Title: Transdiagnostic latent factors dissociating depression and anxiety through reinforcement learning

Article References: Huang, X., Ni, Y., Wang, Y., Peng, Y., & Li, J. (2026). Transdiagnostic latent factors dissociating depression and anxiety through reinforcement learning. BMC Psychiatry. https://doi.org/10.1186/s12888-026-08592-y

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08592-y

Keywords: reinforcement learning, depression, anxiety, computational psychiatry, learning rate, anhedonia, transdiagnostic, decision-making, mental health, prediction error, BMC Psychiatry, latent factors