The issue turns on a deep divide in measurement theory between what psychometricians call reflective and formative constructs. In a reflective model, an invisible underlying trait—here, a latent quantity of home resources—causes the observable indicators to vary, and the indicators’ mutual correlations determine how much each one counts. In a formative model, the arrows point the other way: the observed components together constitute the composite, and their weights should be justified by what they actually accomplish, such as predicting achievement. Meng and Bolt argue that HRL has always had the hallmarks of a formative index. Books, study supports, parental education, and parental occupation are not interchangeable echoes of a single hidden variable; they are distinct resources that can substitute for one another, and a family can be educationally rich through many different combinations of them.

The dominant practice, however, treats HRL as a latent variable. This tradition brings powerful machinery, including tests of measurement invariance that check whether the indicators relate to the latent construct identically across countries. When those tests fail, researchers typically relax the model through partial invariance, allowing country-specific parameters for the offending items. Wendt and colleagues showed in 2017 that such scaling choices can change estimates of how strongly HRL relates to achievement, and later work on post-conflict societies in the Western Balkans found that the books-at-home item behaved differently across countries, apparently because war and displacement had depleted household libraries independently of other family characteristics. The standard response is to model that anomaly as differential item functioning—a measurement bias to be corrected. Meng and Bolt, drawing on Welzel and Inglehart’s critique and on a 2012 report on socioeconomic status for the US National Assessment of Educational Progress, suggest a provocative alternative: the missing books may not be bias at all, but a genuine resource difference that the index ought to capture.

To test the stakes, the researchers analyzed the 2011 cycle of TIMSS mathematics and PIRLS reading data, covering 185,475 fourth-grade students across 37 countries. Their central tool was a structural equation model in which the five HRL indicators load onto a single latent variable that in turn predicts achievement. If the latent variable faithfully summarizes the achievement-relevant information in its indicators, then each indicator’s factor loading should be roughly proportional to its direct predictive effect on test scores. Any leftover association between an indicator and achievement—detected through modification indices and standardized expected parameter change values—signals a structural violation: the latent variable is over- or under-representing that indicator’s real influence.

The pooled results were striking. Parent-reported number of books in the home carried the strongest loading on the latent HRL variable, about 0.69 for both mathematics and reading, yet its residual effect on achievement was positive and substantial, with expected parameter changes of roughly 0.16 to 0.17. In plain terms, books reported by parents predict achievement considerably more strongly than their loading on the latent variable implies. Parental occupation showed the opposite pattern: despite a healthy loading, its direct relationship with achievement was weaker than the latent model suggested, with negative expected parameter changes around −0.10 to −0.13. The authors offer an intuitive reading. Books in the home reflect not just material resources but a family’s active emphasis on education, whereas parental occupation conveys potential support without saying much about how learning is actually prioritized under the roof.

With the mismatch established, Meng and Bolt compared four ways of building the composite: a latent variable score with weights pooled across countries, a latent variable score with country-specific weights, an achievement-tailored composite with pooled regression weights, and an achievement-tailored composite whose weights come from regressing achievement on the indicators within each country. In the tailored approaches, the regression coefficients themselves become the indicator weights, so the composite is defined by how well its ingredients predict the outcome. Across the full sample, the choice mattered surprisingly little: the various composites correlated at roughly 0.95 or higher, and the average gain in explained achievement variance from tailored weights was only about 0.01 in R-squared. For most countries, the philosophical dispute is nearly academic.

But not everywhere. In a handful of systems, the differences were consequential. In Azerbaijan, the country-specific latent variable and tailored composites correlated at only 0.83, and switching from the latent variable index to the tailored one increased the proportion of explained mathematics achievement variance by about 33 percent. For reading, Norway saw a 44 percent proportional increase in R-squared, from 0.09 to 0.13, when tailored weights replaced country-specific latent variable weights; Ireland gained 22 percent and Saudi Arabia 25 percent. The pattern in Norway and Ireland was the mirror image of good measurement: the latent variable approach, which derives weights purely from how indicators correlate with one another, downweighted the proximal indicators—books and study supports—that actually predict achievement, while upweighting the distal indicators of parental education and occupation. Saudi Arabia displayed its own idiosyncratic profile, with study supports over-weighted and parental education under-weighted by the latent model relative to their true predictive roles.

Perhaps the most counterintuitive finding concerns what happens when researchers do the very thing invariance methodology recommends. Allowing discrimination parameters to vary by country is supposed to improve measurement, yet for many countries it actually reduced the prediction of achievement compared with pooled latent variable weights. The explanation lies in where the weights come from: country-specific loadings are estimated from the internal consistency of the indicators within that country, and there is no guarantee—indeed, no reason to expect—that changes in internal consistency track changes in predictive relevance. A parameter adjustment that makes the measurement model fit better can simultaneously make the composite a worse summary of the achievement-relevant resources a family provides. From a formative perspective, the authors note, partial invariance modeling does not merely fail to fix the problem; it can actively mute meaningful between-country differences by absorbing them into measurement parameters.

The study does not declare a winner. Tailored weights, the authors caution, guarantee better prediction almost by construction, since they are optimized for that purpose—a practice the Cowan report famously warned against, arguing that the relationship between socioeconomic status and achievement should be a finding rather than an optimization exercise. Latent variable weights, by contrast, are defined independently of the outcome and avoid overfitting, but they rest on the assumption that a single hidden cause generates all five indicators, an assumption that recent causal work by VanderWeele shows is frequently violated when indicators have their own direct effects on outcomes. Under such violations, regressing achievement on a latent variable likely underestimates the cumulative influence of the indicators, sometimes substantially.

The broader lesson reaches well beyond the HRL index. International assessments routinely convert background questionnaires into latent constructs—attitudes, motivations, socioeconomic statuses—and then police their comparability with invariance machinery that presumes reflective measurement. Meng and Bolt’s analysis suggests that when a construct is better understood as a bundle of resources than as a hidden trait, that machinery can distort rather than refine cross-country comparisons. The choice between defining equivalence by how indicators relate to each other or by how they relate to an external criterion is not a technical footnote; it changes which countries appear to have strong home-resource effects, by how much, and in which direction. For a field whose rankings move education policy worldwide, the composition of a single index number turns out to be anything but a small detail.

Subject of Research: Methodological comparison of latent variable and achievement-tailored approaches to constructing the Home Resources for Learning index in TIMSS and PIRLS

Article Title: Latent variable or achievement-tailored composite?: The home resources for learning (HRL) index in TIMSS and PIRLS

Article References: Meng, L., & Bolt, D. (2026). Latent variable or achievement-tailored composite?: The home resources for learning (HRL) index in TIMSS and PIRLS. Large-scale Assessments in Education, 14(1), Article 52. https://doi.org/10.1186/s40536-026-00324-8

Image Credits: AI Generated

DOI: 10.1186/s40536-026-00324-8

Keywords: Home Resources for Learning, TIMSS, PIRLS, latent variable models, measurement invariance, formative indicators, reflective indicators, socioeconomic status, structural equation modeling, psychometrics, international large-scale assessments, educational achievement

The core problem with traditional cognitive assessment is statistical. Standard batteries such as the Corsi Block-Tapping task for working memory or the Stroop test for inhibitory control treat each task as an isolated measurement. Researchers typically average across repeated trials — mean response times on the Stroop, for example — and treat trial-to-trial variability as mere noise. Structural equation models can link tasks together, but only under assumptions of linearity that the authors argue oversimplify the true architecture of cognition. The result is a testing paradigm that requires many observations per task before estimates stabilize, and one that fails entirely when a task is skipped or data are missing.

The Washington University team had previously introduced an alternative: the distributional latent variable model, or DLVM. Instead of summarizing performance with a single average, DLVM models the full distribution of an individual’s behavior on each task — capturing both central tendency and variability — and compresses that information into a low-dimensional latent space learned by a neural network. Crucially, every single observation contributes to the estimation of multiple constructs at once, because the model exploits the nonlinear dependencies that link performance across tasks. A person’s position in this learned latent space constitutes their cognitive profile, and the trained model can even generate plausible behavioral data for hypothetical profiles, functioning as a generative oracle for simulation.

Paired with DLVM is a second innovation: distributional active learning, or DALE, a Bayesian algorithm that decides which test item to administer next. DALE frames cognitive testing as sequential Bayesian inference. After each response, it updates a posterior distribution over the individual’s latent position and then selects the next trial by maximizing expected mutual information — in plain terms, it always asks the question that will do the most to shrink its own uncertainty. The algorithm can be primed with a small batch of observations spanning all tasks, in this case two samples per task, before active selection kicks in. This lineage descends from Bayesian active learning methods that have already transformed audiology and vision testing, where adaptive stimulus selection dramatically reduced the trials needed to map perceptual thresholds.

What has been missing until now is ground truth. In the team’s earlier human study, real participants produced real data, but the true underlying cognitive parameters were unknown, making it impossible to measure estimation accuracy precisely. The new paper solves this with an elegant simulation strategy. The researchers trained DLVM models on a retrospective dataset — 88 valid testing sessions in which 18 participants completed up to ten sessions of an eight-task battery over ten days via a mobile app, covering tasks such as the Paced Auditory Serial Addition Test, Countermanding, Running Span, Numerical Stroop, Magnitude Comparison, Corsi Simple and Complex Span, and Cancellation. They then sampled 88 points systematically across the learned latent space and used the model to generate the corresponding ground-truth distributional parameters, from which 240 trial-level observations per task were simulated. Every estimate could now be checked against a known answer.

The first set of analyses pitted DLVM against independent maximum likelihood estimation, or IMLE, the optimal approach if tasks truly were independent. Both models received identical data under equal allocation. The verdict was striking. With only two observations per task, DLVM with two latent dimensions achieved Kullback–Leibler divergence values below 0.200 across all tasks, consistently beating IMLE, with the biggest gains on the sigmoid-shaped span tasks that are notoriously data-hungry. DLVM held its advantage until roughly seven observations per task. More dramatically, in validation analyses DLVM needed only about 20 observations per task — 160 total — to reach near-perfect accuracy in recovering marginal distributions, whereas IMLE required about 100 per task, or 800 total. DLVM could even estimate parameters for tasks that were never administered, something IMLE fundamentally cannot do.

The second stage asked how the way data are collected changes the picture. The team compared six configurations: DLVM or IMLE, each fed by DALE’s adaptive sampling, uniform random sampling, or a traditional fixed test battery delivered in sequential blocks. DALE combined with DLVM was the clear winner in the sparse-data regime, driving KLD below 0.05 by roughly 80 observations. The adaptive algorithm concentrated trials on the complex distributional tasks that carried the most information while allocating fewer trials to simpler accuracy-based tasks, and each simulated session received its own unique, personalized battery. The contrast with conventional practice was stark: at 80 total observations, when a fixed battery had covered only three tasks, DLVM with the battery achieved a KLD of 0.148 while IMLE with the same battery sat at a catastrophic 39.8, simply because it could not infer anything about tasks it had not yet reached.

The authors were careful to probe their own assumptions. Because the simulated data were generated from a DLVM-learned latent space, DLVM might enjoy a structural home-field advantage. So they repeated the analysis using an IMLE-based generative process instead. The qualitative pattern held: IMLE performed best when recovering parameters from data generated under its own specification with abundant data, but DLVM and DALE retained their advantage whenever data were sparse. The team also examined DALE’s trajectories through latent space, finding that the algorithm made large corrections in the first trials, converged to a localized region after about 30 observations, and reliably landed in regions of high probability — even when those regions did not coincide exactly with the true latent position, a consequence of the nonlinear latent space admitting multiple equally plausible solutions. Mean root mean squared error across all 88 sessions was 1.02, and only seven sessions converged to positions with normalized negative log probability above 0.05.

Perhaps the most counterintuitive insight concerns the trade-off between model flexibility and data hunger. In most of machine learning, highly flexible models like deep neural networks need enormous datasets to converge. Here the logic inverts: IMLE, the more flexible estimator, only overtakes DLVM once more than 800 observations are available under these testing conditions, while DLVM’s deliberately constrained low-dimensional embedding extracts meaningful inference from a handful of trials. The authors even observed that random sampling eventually surpassed active learning at very large sample counts, suggesting that switching to random sampling once DALE plateaus — or refining its acquisition function — could reveal additional structure in the data. Both of the study’s pre-registered hypotheses were supported by the results.

The practical implications are considerable. DALE could support adaptive assessment in clinical screening, longitudinal monitoring of cognitive change, and large-scale educational studies where testing time is limited and participant burden matters. One examinee might receive extra Stroop trials while another gets more span or countermanding items, depending on where their performance remains most uncertain, and the output is an uncertainty-aware summary of observable performance rather than a brittle point estimate. The authors argue that future behavioral tasks should be designed to be multidimensional and fully featured so that adaptive algorithms never need to repeat an item, since unsampled regions of a feature space are almost always more informative than sampled ones. For a field long anchored to rigid, hour-long batteries, the message is clear: cognition can be measured faster, smarter, and more personally than ever before.

Subject of Research: Bayesian distributional latent variable modeling and adaptive active learning for efficient assessment of executive functioning

Article Title: Bayesian distributional models of executive functioning

Article References: Kasumba, R., Lu, Z., Marticorena, D. C. P., Zhong, M., Beggs, P., Pahor, A., Ramani, G., Goffney, I., Jaeggi, S. M., Seitz, A. R., Gardner, J. R., & Barbour, D. L. (2026). Bayesian distributional models of executive functioning. Behavior Research Methods, 58(11), Article 310. https://doi.org/10.3758/s13428-026-03191-x

Image Credits: AI Generated

DOI: 10.3758/s13428-026-03191-x

Keywords: executive function, Bayesian inference, active learning, latent variable models, machine learning, cognitive assessment, working memory, inhibitory control, cognitive flexibility, psychometrics, neural networks, adaptive testing

The intellectual foundation of the new work lies in the concept of neural manifolds. When neuroscientists record from many neurons at once, they find that the activity of the population does not explore all of the dimensions theoretically available to it. Instead, the patterns of co-activation are confined to a lower-dimensional subspace, a manifold whose axes reflect the correlated structure of the circuit. Pioneering experiments by Sadtler and colleagues in 2014, using brain-computer interfaces in monkeys, demonstrated that animals could readily learn to control a cursor when the mapping between neural activity and cursor movement stayed within the intrinsic manifold of their motor cortex. When the mapping was altered to require activity patterns outside that manifold, performance collapsed and learning proceeded far more slowly. Subsequent work by Oby and colleagues in 2019 and by Golub and colleagues in 2018 extended this principle, showing that the boundaries of the manifold impose real, measurable constraints on what a brain can learn to do.

Those experiments, however, were performed with invasive recordings in animal models. Whether the same geometric principle governs learning in the intact, behaving human brain remained an open question, and a technically daunting one. Human learning is typically studied with functional magnetic resonance imaging, which measures blood-oxygen-level-dependent signals across the cortex rather than the firing of individual neurons. Busch and colleagues tackled the challenge by extracting the latent structure of whole-brain activity from imaging data, characterizing the manifold of human brain activity at a population level and then asking how that structure changes, or fails to change, as people learn new tasks. Their central result is elegantly asymmetric: learning proceeds smoothly when the required new activity patterns remain within the existing manifold, but is sharply limited when the task demands patterns that lie beyond it.

The technical machinery behind this conclusion deserves attention, because it represents a methodological bridge between two levels of analysis that have long been difficult to reconcile. Dimensionality reduction techniques, such as principal component analysis and related latent variable methods, allow researchers to compress the activity of thousands of voxels into a small number of dominant activity patterns. The geometry of these patterns, their relative positions, orientations and curvatures in the reduced space, constitutes a fingerprint of the brain’s intrinsic organization. By tracking how task-related activity trajectories move through this space during training, the researchers could quantify precisely how far a learner had to push their neural activity away from its habitual configuration. Distance within the manifold predicted rapid improvement; distance beyond the manifold predicted stagnation. The geometry was not a passive byproduct of learning. It was the variable that best explained the pace of learning itself.

This asymmetry carries profound implications for how neuroscientists think about plasticity. The traditional view of learning emphasizes synaptic modification: connections strengthen or weaken, and behavior changes accordingly. The manifold framework does not contradict that view, but it adds a crucial layer of constraint. Synaptic plasticity may be abundant and readily available, yet if the circuit’s correlated structure funnels activity along a limited set of axes, then only certain behavioral changes are cheap to produce. Changing the manifold itself, reorganizing the fundamental correlations that define it, appears to be a far more expensive operation for the brain. In effect, the brain possesses two distinct currencies of adaptation: cheap movements within its existing neural geometry, and costly renovations of that geometry. Busch and colleagues provide the strongest human evidence yet that the exchange rate between these currencies governs what we can learn and how fast.

The findings also resonate with an older literature on muscle synergies and motor primitives. Work by Berger, Gentner, Edmunds, Pai and d’Avella in 2013 showed that the motor system composes movements from a small repertoire of coordinated activation patterns rather than controlling every muscle independently. If behavior is built from a finite set of building blocks, then learning new behaviors is easiest when the desired behavior can be assembled from blocks already in stock. The manifold results generalize this intuition from muscles to cortex: the brain’s activity repertoire constrains the behaviors it can readily acquire. Interestingly, the conceptual reach of neural geometry now extends beyond neuroscience altogether. Research in artificial intelligence, notably the work of Elhage and colleagues on features and representations inside neural networks, has revealed that artificial deep networks also organize their computations along low-dimensional manifolds, and that the geometry of these representations shapes what the networks can learn. Brains and machines, it seems, confront the same geometric trade-offs.

For the growing field of brain-computer interfaces, the implications are immediate and practical. A user learning to control a robotic limb or a communication prosthesis is, in geometric terms, being asked to steer their neural activity into configurations that map onto device commands. The new results suggest that interface designers should work with the brain’s intrinsic geometry rather than against it, aligning decoder mappings with the existing manifold to shorten training times and boost performance. Conversely, when a therapeutic goal requires the brain to adopt genuinely novel activity patterns, clinicians may need to design training protocols that gradually reshape the manifold itself, perhaps through carefully staged practice that pulls the geometry outward over time. Similar reasoning applies to neurorehabilitation after stroke, where recovery may depend on whether the damaged cortex can re-express lost skills within its surviving manifold or must construct new geometric structure from the ground up.

The study also sharpens a question that has lingered since the first manifold experiments: what determines the geometry in the first place? Candidates include the statistics of a lifetime of experience, the anatomical wiring of cortical circuits, developmental programs and genetic constraints. If intrinsic geometry is largely fixed by anatomy and history, then the fastest route to enhanced learning may be to choose training tasks that sit comfortably within a person’s existing manifold, a principle that could personalize education and skill training. If, on the other hand, geometry is malleable with the right kind of experience, then identifying the activities that expand manifolds becomes a central goal for learning science. Busch and colleagues’ human imaging paradigm provides a tool for testing these possibilities, because it allows researchers to measure manifold structure repeatedly as individuals accumulate different kinds of experience.

Writing in Nature Neuroscience, Batista emphasizes the elegance of the asymmetry at the heart of the new findings: the same geometric framework that explains effortless learning also explains its limits, unifying what previously looked like separate phenomena under a single principle. The work transforms a philosophical question, why some skills come easily and others do not, into a quantitative one, measurable in the coordinates of brain activity. As imaging methods and latent variable models continue to improve, researchers may soon be able to read out an individual’s neural geometry and predict, before the first practice session, which skills that person will master quickly and which will demand a renovation of the brain’s deepest structure. Learning, on this view, is not a climb up an abstract ladder of difficulty. It is a journey through a landscape whose contours were drawn long before the journey began, and knowing the map may be the first step toward redrawing it.

Subject of Research: How the intrinsic geometry of human brain activity constrains and guides learning

Article Title: Neural geometry guides learning

Article References: Batista, A. P. (2026). Neural geometry guides learning. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02442-6

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02442-6

Keywords: neural manifolds, neural geometry, motor learning, functional MRI, brain-computer interfaces, neural dimensionality, cortical plasticity, skill acquisition, latent variable models, human neuroscience, Neural, geometry