Monica D. Rosenberg

Neural encoding modeling for narrative comprehension—using LLM embeddings to predict the temporal dynamics of neural activation—is widely used to evaluate how well the brain represents certain semantic information. This approach, however, typically collapses semantic representation into a scalar fidelity score aggregated across participants, overlooking potentially rich individual differences and qualitative differences in naturalistic neural representations. In the current study, we addressed this gap by investigating how neural semantic encoding varies across individuals with sustained attention, and whether this difference is qualitative in addition to quantitative. Using fMRI data from 49 participants listening to an audiobook, Le Petit Prince , in English, we trained encoding models with GPT-2 embeddings at varying layers and forecasting lengths. Encoding models trained on high-level embeddings (from the 8th layer of GPT-2) showed significant brain-LLM alignment in 73 of 122 cortical and subcortical ROIs, concentrated in the temporal cortex and the default mode network. Functional connectivity strength in a validated connectome-based predictive model of sustained attention correlated with encoding performance in 18 ROIs, most strongly in the insula and in control and attention networks, with no ROI showing significant opposite relationship. Decomposing semantic information by levels of abstraction showed that the neural signature of sustained attention was selectively associated with a greater proportion of high-level gist information encoded in the frontoparietal and attention networks, while the opposite relationship emerged for proportion of low-level information. Sustained attention scores further predicted stronger predictive coding of upcoming narrative content across multiple networks across the whole brain (control, default mode, salience/ventral attention, limbic, and somatomotor networks). These findings suggest that sustained attention does not uniformly amplify neural semantic encoding, but instead shifts neural representations toward more compressed, gist-level formats that support anticipation of upcoming content. Individual differences in brain-LLM alignment can be a signature of representational strategy rather than noise, with potential utility for cognitive phenotyping.

Attention prioritizes relevant information from the environment, but does not do so consistently over time. Instead, we sometimes experience poor attentional states which are characterized by a higher degree of performance failures, or "lapses." The source of these lapses--whether they arise from attentional disengagement, failures of attentional selection, or simply failures of response control--is not clear. This study tests whether the strength of neural representations differs as a function of both attentional selection and modulation over time. We recorded electroencephalography (EEG) data while participants performed a sustained attention task requiring the selection of task-relevant items and withholding of responses to rare targets. Results show that relevant items receive prioritized processing in brain signals, as measured with EEG decoding. Yet, their neural representational fidelity weakens under poor sustained attentional states, foreshadowing upcoming behavioral failures. These findings support a neural basis of disengagement during poor attentional states, suggesting that lapses may arise from a decoupling of brain activity from task processing. Finally, pre-item representational strength predicted memory success for targets, suggesting that neural fidelity indexes an encoding-ready state. Results suggest that the quality of neural representations is affected by ongoing attention and plays a key role in processing and memory.

Memory encoding must accommodate increasingly demanding loads, yet it remains unclear whether the underlying brain state is preserved or adaptively reconfigured as load increases. Here, we recorded EEG from 111 participants performing a visual memory task in which they encoded lists of real images ranging from 1 to 128 items, spanning low to high mnemonic demands, followed by a recognition test. We measured the similarity of set-size-specific interelectrode correlation patterns to a low-load reference. Similarity decreased monotonically as list length increased: patterns at set sizes below 4 remained close to the low-load configuration, whereas set sizes of 8 and above progressively shifted toward a distinct high-load configuration. This transition was captured by a single rotating eigenvector, revealing a low-dimensional trajectory from low- to high-load neural geometry. Smaller rotation angles from the low-load template predicted better memory performance at high loads. Together, these findings provide evidence for adaptive reconfiguration and reveal a structured, low-dimensional neural trajectory through which encoding architecture changes as mnemonic demands increase.

Alcohol drinking motives vary among individuals and shape experiences and beliefs about alcohol, influencing the processing of alcohol-related cues. In real-life settings, these cues are contextually rich, amplifying the role of such individualized drinking motives on cue processing. However, previous literature has primarily relied on images of alcohol, which lack contexts and differ significantly from real-life. Here, aiming to investigate real-life craving, we examined the role of alcohol drinking motives in craving in response to naturalistic alcohol-drinking videos. We asked fifty-three problematic alcohol users to speak about their reasons for drinking alcohol to capture unique alcohol drinking motives of each individual. Participants also underwent functional MRI while watching fifteen alcohol-drinking videos, and reported their subjective level of craving and self-relatedness for each video. Behavioral data analysis revealed that individuals with greater alcohol use severity tended to report greater cue-induced craving, but only when they reported that a video was related to themselves. Inter-subject representational similarity analysis showed that participants with similar alcohol drinking motives, reflected in shared drinking reasons and similar self-relatedness to the videos, exhibited synchronized craving-related neural responses during video-watching. Notably, these shared neural processes mediated the link between similar drinking motives and similar self-reported craving levels across participants. Together, our findings highlight the crucial role of alcohol drinking motives in shaping cue-induced alcohol craving, and provide deeper insights into craving in real-world contexts.

We make sense of everyday events by reasoning about their underlying causes. When we connect causal links between events separated in time, we often experience a sudden feeling of "aha!", or insight. To understand its cognitive and neural mechanisms, we designed an fMRI study in which participants watched a temporally scrambled TV episode. Participants pressed an "aha" button whenever they understood something new and explained their reasoning. More than 40% of insight explanations reference past events causally related to the current event. Neural patterns representing these causally related past events are reinstated across the cortex. This neural reinstatement drives sudden shifts in cortical activity patterns ~2 s prior to aha button presses, reflecting updates in situational representation. Moreover, distributed brain areas represent causally related events with similar neural patterns, beyond shared semantic or perceptual features. Together, the study suggests that we comprehend events by retrieving causally related past events, followed by updating neural patterns at moments of insight.

Mind-wandering (MW) is a frequent and pervasive phenomenon, yet it is commonly assessed using self-reports or probe-based methods that offer limited temporal precision regarding its onset. In this study, we introduce a novel paradigm, ReMind, that estimates the onset and duration of MW episodes during natural reading by combining retrospective self-reports with eye-tracking. Participants indicated the words where they believed their mind started and stopped wandering, and these reports were aligned with gaze timestamps to estimate MW onset. Using data from 44 participants, we examined whether knowledge of MW onset improves the detection of MW from eye-tracking signals. To evaluate relevance for both self-report and thought-probe paradigms, we additionally simulated thought probes by randomly sampling time points during reading. Logistic regression classifiers trained on eye-tracking features extracted from time windows anchored to MW onset achieved AUROC scores of 0.659 and 0.621 under the self-report and simulated thought-probe paradigms, respectively, using leave-one-subject-out cross-validation. In both cases, onset-aligned windows outperformed classifiers trained using arbitrary MW windows. Sliding-window analyses further revealed systematic temporal changes around MW onset, with classification performance peaking at approximately 3 seconds after onset. Feature-level analyses showed reduced fixation rate and fixation dispersion, along with increased pupil size following MW onset. Together, these findings characterize the temporal progression from on-task reading to MW. Overall, ReMind provides a useful framework for studying the temporal dynamics of MW during naturalistic reading.

Updating impressions of others is essential to navigating social life. As we get to know an individual, we update our impressions of them in accordance with new information. How the brain dynamically revises impressions in real time under naturalistic conditions remains unclear. Here, we address this question using functional magnetic resonance imaging (fMRI) and natural language analysis in a naturalistic social cognition paradigm. Across 10 runs, participants viewed a character-driven TV episode, reported moments of insight, and described their impressions of the characters. Results reveal that impressions progressively evolved over time. Individuals with more similar existing impressions exhibit greater neural synchrony during movie-watching, which predicts convergence in post-movie impressions. Neural synchrony in the right superior temporal sulcus (STS) mediates the influence of initial similarity on later alignment. Insight moments accompany neural pattern shifts widespread across cortical regions, including the temporoparietal junction (TPJ), dorsomedial prefrontal cortex (dmPFC) and STS, and the magnitude of the shifts tracks the degree of impression updating. Specifically, distinct forms of insight selectively update complementary components of impressions: character insight updates person-centered representations, whereas non-character insight shapes social-event structure. Together, these findings show that people update their impressions of others through dynamic shifts in distributed brain activity patterns at moments of insight, providing a novel and ecologically grounded neural account of social cognition. Significance statement: People do not simply form first impressions and keep them. As we learn new things about others, our judgments change—sometimes gradually, sometimes in a flash of insight. By tracking brain activity while people watched a TV show and described how their impressions of the characters evolved, we found that people who saw a character in more similar ways also show more similar brain responses when watching them onscreen, which in turn leads to convergence in subsequent impressions. Such updating is closely tied to “aha” moments, when a person is suddenly understood in a new way. These moments were marked by rapid shifts in the brain representational patterns, and larger shifts predicted greater changes in impressions. These findings offer a more naturalistic account of how the brain revises our understanding of other people in everyday life.

Classic memory models proposed that the encoding process involved in visual working memory (VWM) controls the bandwidth of encoding in long-term memory (LTM). Behaviorally, VWM and LTM accuracies are reliably correlated at the behavioral level, raising the question of whether LTM encoding uniquely engages processes that are distinct from VWM encoding. To investigate this, we recorded EEG activity as participants completed recognition memory tasks with set sizes of 32 and 128, far beyond typical VWM capacity. Using interelectrode correlation (IC) analysis, we found that IC patterns reliably predicted individual differences in LTM encoding across both set sizes, indicating a robust, domain-general neural signature. Importantly, this predictive power remained even after controlling for VWM and attentional control performance, suggesting that the model captures variance specific to LTM encoding. Temporally, predictive signals emerged only after stimulus onset and persisted for 500–600 ms. Early and late encoding phases involved distinct network structures, reflecting dynamic neural processes underlying individual differences in LTM encoding. Lastly, we showed that alpha band-passed IC, but not theta or beta band-passed IC, selectively predicted individual differences in LTM performance. Together, our findings reveal a unique and temporally dynamic neural signature that supports individual differences in LTM encoding, independent of general cognitive abilities.

Autism is a heterogeneous condition, and functional magnetic resonance imaging-based studies have advanced understanding of neurobiological correlates of autistic features. Little work has focused on the optimal brain states to reveal brain–phenotype relationships. Here, using connectome-based predictive modeling, we interrogated four datasets to determine scanning conditions that boost prediction of clinically relevant phenotypes and assess generalizability. In dataset one, a sample of youth with autism and neurotypical participants (n = 63), we found that a sustained attention task resulted in high prediction performance of autistic traits compared with a free-viewing social attention task and a resting-state condition. In dataset two (n = 25), we observed the predictive network model of autistic traits generated from the sustained attention task generalized to predict measures of attention in neurotypical adults. In datasets three and four, we determined the same predictive network model further generalized to predict measures of social responsiveness in the Autism Brain Imaging Data Exchange (n = 229) and the Healthy Brain Network (n = 643). Our data suggest an in-scanner sustained attention challenge can help delineate robust markers of autistic traits.