Raymond J. Dolan

A neural mechanism for compositional structure transfer in humans

A human ability to adapt to the dynamics of novel environments relies on abstracting and generalizing from past experiences. Previous research has focused on how humans generalize from isolated sequential processes, yet we know little about mechanisms that enable adaptation to more complex dynamics, including those that govern much everyday experience. Here, using a novel sequence learning task based on graph factorization, coupled with simultaneous magnetoencephalography (MEG) recordings, we ask whether reuse of experiential “building blocks” enable inference and generalization. Behavioral results were consistent with participants decomposing task experience into subprocesses, abstracting their dynamical structure away from their sensory specifics and transferring these to a new task environment. Neurally, this transfer was associated with a learning-induced, condition-specific, increase in neural similarity among stimuli conditional on their structural role in abstract subprocesses that were shared across task phases. Consistent with this, a complementary dynamical role decoding analysis revealed enhanced neural decodability for stimuli sharing identical structural transition types across experiential contexts, specifically when prior experience included the relevant graph structure. Decoding strength for these role representations was positively related to behavioral success in subprocess knowledge transfer, a relationship that will require future independent replication. These findings provide neural evidence consistent with the idea that a structural scaffolding mechanism supports generalization of experience to new contexts.

Distinguishing learning asymmetry from perseveration in behavioral models

A central challenge in cognitive science is to distinguish between multiple processes that can result in similar behaviors. In reinforcement learning, one prominent example concerns two potential drivers of choice repetition: a learning asymmetry, in which agents learn differently from outcomes depending on their valence and/or whether they confirm choice, and choice perseveration, an outcome-independent tendency to repeat past choices. Evidence for asymmetric learning has typically relied on computational models that control for perseveration, or specific behavioral markers. Here, we show both these approaches have critical flaws and can spuriously detect learning asymmetries even in perseverative, symmetric-learning agents. To address this, we introduce a statistical test that distinguishes genuine learning asymmetries from spurious effects. Applying this test to a large dataset spanning ten published experiments across four studies, we find that some previously reported confirmation biases are fragile, albeit others remain robust even at an across-study level. Finally, we propose a task design that can yield a more valid qualitative signature of confirmation bias. Our approach provides a reliable framework for disentangling processes underlying choice repetition, while providing tools for the wider research community that can minimize potential spurious effects arising from process mimicry and biased parameter estimation.

Digital episodic memory tools for the identification of amnestic mild cognitive impairment: a systematic review and an exploratory meta-analysis

Systematic review demonstrates digital episodic memory tools improve aMCI diagnosis, suggesting scalable options for evaluation.

Trait-like visual cortical hyperactivity in trait anxiety

Sensory processing varies across individuals, with some traits-particularly sensory hypersensitivity to basic non-valenced stimuli-linked to emotional traits and psychiatric risk. Traditional accounts attribute this sensory-emotion linkage to limbic or prefrontal modulation, but empirical support is limited. Growing evidence suggests sensory cortex itself flexibly encodes value beyond labeled-line analysis. Across four high-density EEG experiments with multi-wave assessments, we identified reliable visual cortical hyperactivity in high trait anxiety. The effect emerged as early as 46 ms, localized to V1/V2, and was specific to the parvocellular pathway. It was reproducible across arousal states, stimulus valence, extended intervals, and paradigms, and evident for both simple (grating) and complex real-world images. Importantly, cortical excitation-inhibition balance (EEG aperiodic exponent/1/f slope) predicted parvocellular responses in low- but not high-anxiety individuals, implicating disrupted E/I modulation. Thus, trait anxiety alters early visual processing, aligning cortical computations with an individual's intrinsic biological propensity from the outset.

Anxious vision: Trait-like visual cortical hyperactivity in trait anxiety

Sensory processing varies across individuals, often manifesting as distinct traits. Some traits, particularly sensory hypersensitivity, even to basic non-valenced stimuli, are closely linked to emotional traits and implicated in a diversity of psychiatric disorders. Extant theories attribute this sensory-emotion linkage to limbic and prefrontal modulation, but empirical support remains limited. Meanwhile, it is increasingly recognized that sensory cortex supports flexible, value-based processing beyond labeled-line sensory analysis. Assaying early visual cortical activity (using visual evoked potentials/VEPs) across four high-density electroencephalogram (hd-EEG) experiments, involving three independent samples ( N = 150) and repeated assessments (up to three waves), we identify a reliable and generalizable pattern of visual cortical hyperactivity related to high trait anxiety. Notably, this effect was specific to the parvocellular (vs. magnocellular) pathway, emerging as early as 46 ms and localized to V1. Underscoring its trait-like stability, the effect was reproducible across different arousal states, neutral and negative stimuli, extended time intervals, and varied experimental paradigms. Moreover, its ecological validity was affirmed by its presence in response to both basic low-level stimuli (Gabor patches) and complex real-world images. Crucially, an excitation-inhibition (E/I) ratio (indexed by the resting-state EEG aperiodic exponent; 1/f slope) predicted parvocellular VEP magnitude in subjects with low but not high trait anxiety, suggesting disrupted E/I modulation in trait anxiety underpins this hyperactivity. We conclude that in trait anxiety, early visual processing deviates from strict “fidelity” to physical-world input or mere “subservience” to top-down modulation, instead adapting to an individual’s biological propensity at the earliest processing stage. Highlights Hyperactivity in primary visual cortex is stable, trait-like It occurs as early as 46 ms post-stimulus, during initial feedforward sweep It covaries with trait anxiety, a key personality factor and risk predisposition for psychiatric disorders It is related to disrupted excitation-inhibition modulation of visual cortex in trait anxiety

Habenula–ventral tegmental area functional coupling and risk aversion in humans

Maladaptive responses to uncertainty, including excessive risk seeking or avoidance, are linked to a range of mental disorders. One expression of these is a provariance bias (PVB), i.e., risk-seeking manifests as a preference for choosing options with higher variances/uncertainty. Using a magnitude learning task, we show that individual differences in PVB are explained by a model that includes asymmetric learning rates, allowing differential learning from positive prediction errors (PPEs) and negative prediction errors (NPEs). Using high-resolution 7T functional MRI (fMRI), we identify distinct neural responses to PPEs and NPEs in value-sensitive regions, including habenula (Hb), ventral tegmental area (VTA), nucleus accumbens (NAcc), and ventral medial prefrontal cortex (vmPFC) in humans. Notably, prediction error signals in NAcc and vmPFC were boosted for high variance options. NPEs responses in NAcc were associated with a negative bias in learning rates that was linked to the strength of Hb-VTA negative functional coupling during NPE encoding, with a mediation analysis revealing this coupling influenced NAcc responses to NPEs via an impact on learning rate bias. Our findings implicate Hb-VTA functional coupling in the emergence of risk preferences during learning, with implications for psychopathology.

Characterizing the Paradoxical Interplay of Semantic Perseveration and Derailment in Psychiatry Using Theory-Driven Generative Language Simulations

BACKGROUND AND HYPOTHESIS: Formal thought disorder (FTD), characterized by derailment and perseveration, is a central feature in psychosis and can partially appear in other psychiatric conditions. Despite their clinical co-occurrence, derailed speech and perseveration have opposing effects on common natural language processing (NLP) metrics of FTD (eg, cosine distance metrics). We propose a set of novel NLP metrics suitable for capturing perseveration, separate from any degree of (dis)organization and derailment. STUDY DESIGN: We generate a synthetic dataset using large language models to simulate narratives while independently manipulating semantic perseveration (via paraphrase insertion during generation) and derailment (by increasing word selection stochasticity). We developed novel metrics aiming to capture perseveration by probing semantic structure and density (using clustering and dimensionality reduction). These metrics were compared against traditional cosine distance metrics in the synthetic dataset, as well as in a transdiagnostic general population sample (n = 811). STUDY RESULTS: Traditional cosine distance metrics performed well in detecting derailment but underperformed for perseveration. Conversely, density-based metrics showed superior performance in detecting perseveration and dissociating it from derailment. Density metrics also outperformed traditional metrics in revealing perseveration in psychopathological dimensions associated with repetitive thinking and speech, including depressive symptoms and difficulty initiating conversation (ie, negative FTD). CONCLUSIONS: Integrating density-based metrics with traditional semantic distance measures may better capture FTD symptoms, though further validation in clinical samples is needed. This approach improves understanding of perseveration and its interaction with (dis)organization in thought and speech, with potential clinical applications.

Disinformation elicits learning biases

Experiments reveal how disinformation distorts learning processes, suggesting new strategies for mitigation.

Learning asymmetry or perseveration? A critical re-evaluation and solution to a pervasive confound

This research introduces a new statistical test for assessing learning asymmetries in reinforcement learning, suggesting implications for future studies.

Investigating the Role of the Locus Coeruleus Noradrenergic System in Cognitive Function in Amnestic Mild Cognitive Impairment: An fMRI Study

fMRI reveals links between locus coeruleus dysfunction and cognitive decline in individuals with amnestic mild cognitive impairment, indicating potential benefits of noradrenergic interventions.

Challenges in Replay Detection by TDLM in Post-Encoding Resting State

Analysis reveals lack of replay detection during resting state in humans, indicating issues with statistical power.

Trait anxiety is associated with reduced reward-related replay at rest

Differences in LC integrity and fMRI activations in healthy aging and MCI associated with successful memory encoding

Observational analysis shows reduced LC integrity impacts encoding success in MCI and older adults, suggesting neural mechanisms for memory decline.

The Semantic Underpinnings of Speech Disorganization in Schizophrenia

Analysis using magnetoencephalography reveals unique semantic structures in schizophrenia, suggesting personalized models improve understanding of speech disorganization.

Profiling brain morphology for autism spectrum disorder with two cross-culture large-scale consortia

We explore neurodevelopmental heterogeneity in Autism Spectrum Disorder (ASD) through normative modeling of cross-cultural cohorts. By leveraging large-scale datasets from Autism Brain Imaging Data Exchange (ABIDE) and China Autism Brain Imaging Consortium (CABIC), our model identifies two ASD subgroups with distinct brain morphological abnormalities: subgroup “L” is characterized by generally smaller brain region volumes and higher rates of abnormality, while subgroup “H” exhibits larger volumes with less pronounced deviations in specific areas. Key areas, such as the isthmus cingulate and transverse temporal gyrus, were identified as critical for subgroup differentiation and ASD trait correlations. In subgroup H, the regional volume of the isthmus cingulate cortex showed a direct correlation with individuals’ autistic mannerisms, potentially corresponding to its slower post-peak volumetric declines during development. These findings offer insights into the biological mechanisms underlying ASD and support the advancement of subgroup-driven precision clinical practices. Using the lifespan brain chart and two large-scale cross-cultural consortia (ABIDE and CABIC), we disentangle the heterogeneity of brain morphology in ASD into two structural impairment pathways that could disrupt sensory to association functions.

Pupil dilation as a marker of attention/effort in aging and mild cognitive impairment

This research demonstrates pupil dilation reflects attention and cognitive effort in aging and mild cognitive impairment, suggesting potential as a cognitive assessment tool.

A neural mechanism for compositional generalization of structure in humans

Experiments reveal how neural mechanisms adaptively utilize subprocesses for generalization in novel environments, implying efficiencies in learning.

When Collaboration Falters, Insensitivity to How Our Actions Affect Others Drives Inflated Self-evaluations

During high-stake interactions, people not only evaluate policies or outcomes, but also themselves and others. Such evaluations may be crucial for long-term outcomes, such as harmonious marriage, confident leadership and indeed mental health. Powerful evaluations occur during interactions, where people can support or let each other down. Thus, we implemented an interactive decision-making game, wherein two real-life participants explicitly evaluated themselves and their play-partner while playing an ecologically framed, probabilistic, iterated prisoner's dilemma. To separate preferences from abilities, participants did not interact with the other directly, but instructed a computer avatar on how to play on their behalf. We tested a range of computational models of participants' person-evaluations. In some, self-evaluation relied on regret or satisfaction regarding one's decisions. However, the winning models relied directly on observed gains and losses. Here, evaluation of the self was proportional to how much one's partner benefited, and vice versa. We found a marked self-positivity bias, which was most prominent in dyads where both partners often defected. Between participants, a self-positivity bias was explained by a reduced weight of one's partner's benefits onto self-evaluation. This suggests that the negative outcomes claimed to attract defensive, external attribution by attribution theorists are one's partner's poor outcomes. Further analysis suggested that a reduced sensitivity to others' outcomes was associated with reduced earnings for the self, hinting at a functional role for person-evaluations in decision-making. Thus, we introduce a novel computational model that provides a concise account of self-serving bias in evaluations, as observed during risky dyadic interactions. Supplementary Information: The online version contains supplementary material available at 10.1007/s42113-025-00250-y.

Challenges in Replay Detection by TDLM in Post-Encoding Resting State

Using temporally delayed linear modeling (TDLM) and magnetoencephalography (MEG), we investigated whether items associated with an underlying graph structure are replayed during a post-learning resting state. In these same data, we previously provided evidence for replay during online (non-rest) memory retrieval. Despite successful decoding of brain activity during a localizer task, and contrary to predictions, we found no evidence for replay during a post-learning resting state. To better understand this, we performed a hybrid simulation analysis in which we inserted synthetic replay events into a control resting state recorded prior to the actual experiment. This simulation revealed that replay detection using our current pipeline requires an extremely high replay density to reach significance (>1 replay sequence per second, with ‘replay’ defined as a sequence of reactivations within a certain time lag). Furthermore, when scaling the number of replay events with a behavioral measure, we were unable to induce a strong correlation between sequenceness and this measure. We infer that even if replay was present at plausible rates in our resting state dataset, we would lack statistical power to detect it with TDLM. Finally, contrasting our novel hybrid simulation to existing purely synthetic simulations indicated that the latter approaches overestimate the sensitivity of TDLM. We discuss approaches that might optimize the analytic methodology, including identifying boundary conditions under which TDLM can be expected to detect replay. We conclude that solving these methodological constraints will be crucial for optimizing the non-invasive measurement of human replay using MEG.

Disinformation elicits learning biases

In open societies disinformation is often considered a threat to the very fabric of democracy. However, we know little about how disinformation exerts its impact, especially its influences on individual learning processes. Guided by the notion that disinformation exerts its pernicious effects by capitalizing on learning biases, we ask which aspects of learning from potential disinformation align with ideal “Bayesian” principles, and which exhibit biases deviating from these standards. To this end, we harnessed a reinforcement learning framework, offering computationally tractable models capable of estimating latent aspects of a learning process as well as identifying biases in learning. In two experiments, participants completed a two-armed bandit task, where they repeatedly chose between two lotteries and received outcome-feedback from sources of varying credibility, who occasionally disseminated disinformation by lying about true choice outcome (e.g., reporting non reward when a reward was truly earned or vice versa). Computational modelling indicated that learning increased in tandem with source credibility, consistent with ideal Bayesian principles. However, we also observed striking biases reflecting divergence from idealized Bayesian learning patterns. Notably, in one experiment individuals learned from sources that should have been ignored, as these were known to be fully unreliable. Additionally, the presence of disinformation elicited exaggerated learning from trustworthy information (akin to jumping to conclusions) and exacerbated a normalized measure of “positivity bias” whereby individuals self-servingly boost their learning from positive, relative to negative, choice-feedback. Thus, in the face of disinformation we identify specific cognitive mechanisms underlying learning biases, with potential implications for societal strategies aimed at mitigating its harmful impacts.