The work, led by Sahar Khosravi and Behtash Babadi of the University of Maryland together with Nikolas Francis and Patrick Kanold, introduces what the authors call a Granger sensori-behavioral functional taxonomy, or G-taxonomy, for neuronal ensembles recorded with two-photon calcium imaging. The central idea is deceptively simple: instead of treating encoding, connectivity, and behavioral readout as separate analyses, the framework extracts all of them from the same data in a single coherent model, using the language of Granger causality to describe how information flows between stimuli, neurons, and behavior.

Granger causality, a concept borrowed from econometrics, defines a directional influence in terms of temporal predictability. If the past activity of neuron A improves predictions of neuron B beyond what B’s own past already reveals, then A is said to Granger-cause B. Applied to neuroscience, this notion promises something remarkable: a wiring diagram of functional influence that is estimated directly from data, without requiring the invasive perturbation of individual cells. But applying it to two-photon calcium imaging has proven notoriously difficult, and the new paper confronts those difficulties head-on.

The problem begins with the physics of the measurement itself. Two-photon microscopy does not record spikes, the electrical impulses neurons use to communicate. Instead, it tracks fluorescent calcium indicators whose glow rises and falls on a much slower timescale than the underlying spiking activity. Calcium transients are sluggish, indirect, and corrupted by noise, which means the fast temporal structure that Granger analysis depends on is largely hidden from view. On top of this, the relationship between spikes and calcium is nonlinear, and the recorded signals reflect a mixture of genuine neural dynamics and the latent interplay between external stimuli and endogenous brain processes.

To overcome these obstacles, the researchers built their framework on an integration of several sophisticated statistical techniques. State-space modeling provides a mathematical scaffold for inferring the latent spiking activity that generated the observed calcium traces, treating the unobserved spikes as hidden states that evolve over time. Variational inference, a modern computational approach popular in machine learning, makes the estimation of these hidden states tractable even for large populations of neurons. Point-process models then describe the stochastic timing of spikes in a statistically principled way, allowing the framework to work with the discrete, event-like nature of neural firing rather than pretending spikes are smooth continuous signals.

With this machinery in place, the framework computes Granger causal effects along three distinct axes: from neuron to neuron, capturing functional connectivity within the recorded ensemble; from external stimuli to neurons, capturing sensory encoding; and from neurons to behavior, capturing the readout side of the circuit. Inspired by the intersection information framework, the method goes one step further and identifies neurons that encode specific features of sensory stimuli which are actually relevant to the animal’s behavioral report, a distinction that most encoding analyses ignore entirely.

The result is a taxonomy that sorts neurons into functionally distinct groups based on their sensori-behavioral relevance. Some cells may encode stimulus features without any apparent link to behavior, others may influence their neighbors without strong sensory tuning, and a select subset may sit at the critical junction, carrying information about the stimulus that demonstrably informs the animal’s decision. This classification, the authors suggest, offers a far richer picture of cortical computation than simple tuning curves or correlation-based connectivity maps can provide.

Before trusting the method with real brains, the team validated it on simulated data, where the ground truth was known. These simulation studies revealed significant improvements over existing techniques, suggesting that the framework recovers directional interactions more faithfully than previous approaches that either ignored the indirect nature of calcium measurements or treated encoding and connectivity in isolation. The simulations also demonstrated that the method remains robust in the presence of the noise levels and slow dynamics characteristic of real imaging experiments.

The researchers then applied their framework to experimentally recorded two-photon imaging data from the mouse auditory cortex, area A1, during two behavioral conditions: passive listening to tones and active tone discrimination. In the passive condition, animals simply heard sounds; in the active condition, they had to discriminate between tones and report their choice, allowing the experimenters to link neural activity to both stimulus and behavior on a trial-by-trial basis. The analysis identified distinct groups of cells with diverse sensori-behavioral roles, confirming that even within a single cortical area, neurons occupy strikingly different positions in the flow from sensation to action.

Perhaps most intriguingly, the framework revealed changes in functional connectivity associated with correct versus incorrect behavioral trials. When mice performed the discrimination task accurately, the pattern of Granger influences among neurons differed from trials in which they erred, hinting that the moment-to-moment state of a cortical network, not just its average properties, shapes whether sensory information is successfully transformed into the right decision. Such trial-level differences are exactly the kind of signal that fragmented analyses tend to miss, because they emerge only when encoding, connectivity, and behavior are examined within the same statistical model.

The implications extend well beyond the auditory cortex. Two-photon calcium imaging has become one of the most widely used tools in modern neuroscience, generating massive datasets from visual cortex, hippocampus, frontal areas, and beyond. A general-purpose method for extracting directional, behaviorally relevant structure from such data could reshape how laboratories interpret their recordings, turning raw fluorescence movies into functional maps of information flow. The framework’s data-driven character means it makes few assumptions about what the circuit should look like, letting the statistics speak for themselves.

There are, of course, important caveats that the authors and the broader field continue to weigh. Granger causality inferred from observational data reflects predictive relationships, not necessarily direct physical connections, and calcium imaging still imposes temporal limits that even the best state-space methods cannot fully erase. The framework also requires substantial computational resources, since variational inference over large populations is demanding. Yet the study’s combination of rigorous simulation benchmarks and successful application to real experimental data suggests these hurdles are surmountable, and that the approach can deliver on its promise in practice.

What makes this work resonate beyond its technical contributions is the question it ultimately addresses: how does a distributed population of neurons transform sensory inputs into behaviorally relevant representations? By refusing to split that question into disconnected sub-problems, the G-taxonomy framework offers a principled answer in the form of a single, unified statistical picture, one in which every neuron’s role, from stimulus encoding to influence on neighbors to behavioral readout, can be quantified from the same recording. As imaging technology continues to scale up to ever larger ensembles, tools like this one may become essential for reading the grammar of cortical circuits, and for understanding not just which neurons fire, but why it matters for what the animal does next.

Subject of Research: A unified Granger causal framework for extracting stimulus-to-neuron, neuron-to-neuron, and neuron-to-behavior interactions from two-photon calcium imaging data

Article Title: Granger sensori-behavioral functional taxonomy of neuronal ensemble activity from two-photon calcium imaging data

Article References: Khosravi, S., Francis, N. A., Kanold, P. O., & Babadi, B. (2026). Granger sensori-behavioral functional taxonomy of neuronal ensemble activity from two-photon calcium imaging data. PLOS Computational Biology, 22(9), e1014820. https://doi.org/10.1371/journal.pcbi.1014820

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014820

Keywords: two-photon calcium imaging, Granger causality, functional connectivity, neural encoding, auditory cortex, state-space modeling, variational inference, point processes, mouse behavior, computational neuroscience, neuronal ensembles, sensory discrimination

The researchers recorded the activity of hundreds of individual hippocampal neurons in patients who had electrodes implanted for clinical reasons, typically to localize seizure activity before epilepsy surgery. While the patients listened to narrative speech, the team captured the firing patterns of these neurons with millisecond precision. This rare opportunity to eavesdrop on the living human brain at single-cell resolution allowed the investigators to ask a deceptively simple question: when a person hears a word, what do hippocampal neurons actually do with its meaning?

To answer it, the team turned to an encoding model approach, a technique in which computational models of language are used to predict neural responses. Crucially, the researchers controlled for lower-level features of the speech signal, including the phonemic content of words and their grammatical roles. Even after stripping away these confounds, semantic information remained robustly encoded in the population activity. In other words, the hippocampal neurons were not merely responding to sounds or syntactic structure; they were representing what the words meant.

One of the study’s most intriguing findings concerns the selectivity of individual neurons. Rather than acting as specialized detectors for single words or narrow categories, hippocampal neurons displayed complex, mixed selectivity, responding to multiple words drawn from multiple semantic categories. This kind of distributed, high-dimensional coding has been championed in the population doctrine of cognitive neuroscience, which holds that complex cognitive content is carried not by individual cells but by patterns of activity spread across many neurons. The hippocampal semantic code fits this doctrine well: meaning emerges from the joint pattern of firing across the population, much as a chord carries information that no single note contains.

The parallels with large language models run deeper still. In models such as Word2Vec, BERT and GPT-2, words are represented as vectors in a high-dimensional space, where the distance between vectors reflects semantic similarity. The researchers found that the distance between neural population responses in the hippocampus correlated with semantic distance computed from these embedding spaces. Words with similar meanings evoked population activity patterns that were closer together in the neural space, while unrelated words produced more distant patterns. The geometry of meaning, it seems, is not unique to machines trained on oceans of text; the human hippocampus appears to organize word meanings along comparable geometric lines.

Yet the neural code was not a simple copy of any single model. Among the embedding schemes tested, neural population activity aligned most closely with GPT-2, a contextual transformer model that represents words differently depending on the surrounding sentence. This alignment suggests that hippocampal semantic representations are heavily contextualized: the same word does not always elicit the same neural pattern, but rather a pattern shaped by the linguistic moment in which it appears. Supporting this interpretation, variation in the neural response patterns across occurrences of a word correlated with polysemy measures derived from language models, meaning that words with many context-dependent senses produced more variable neural representations than words with a single, stable meaning.

The study also uncovered a subtle and technically fascinating phenomenon the authors call contrastive coding. For pairs of semantically similar words, the relationship between semantic distance and neural distance inverted: instead of nearby meanings producing nearby neural patterns, the hippocampus pushed similar representations apart. The researchers attribute this counterintuitive pattern to the noise-mitigating benefits of contrastive coding, a principle familiar from machine learning systems that learn by pushing similar examples apart to make them more distinguishable. In the brain, where neural signals are inherently noisy, deliberately separating representations of similar concepts could prevent them from being confused with one another. This echoes earlier findings that overlapping spatial memories in the hippocampus trigger representational repulsion, suggesting the structure may apply a similar differentiation strategy to meaning as it does to space and memory.

These results reshape how neuroscientists think about the hippocampus itself. For decades, the structure has been studied primarily as the engine of episodic memory, the faculty that binds the who, what, where and when of personal experience. But a growing body of evidence, including prior work showing that hippocampal neurons reactivate conceptual representations when pronouns refer back to earlier nouns, and that the region supports the flexible use of language in conversation, points to a broader role. The new findings suggest that representing word meaning during comprehension is not a peripheral sideline for the hippocampus but a core computational function, one that may provide the semantic scaffolding upon which narrative memories are built.

The methodological rigor of the study strengthens its conclusions. The team analyzed responses across thousands of words, validated their models against shuffled controls, and demonstrated that semantic embeddings explained significantly more variance in neural activity than phonetic features alone. They also showed that the semantic-neural distance relationship held across individual patients and that removing function words and pronouns did not eliminate the effect. The data and analysis code have been made publicly available, allowing other researchers to scrutinize and extend the findings, a transparency increasingly expected in high-stakes human neuroscience.

The broader implications reach toward the frontier where neuroscience and artificial intelligence converge. If the human brain and large language models have independently arrived at similar solutions for representing meaning, then the vector-space approach that underpins modern AI may capture something fundamental about how meaning must be organized to be useful. Conversely, the brain’s contrastive tricks and contextual flexibility could inspire the next generation of models. For now, the study delivers a vivid answer to an ancient question: when you hear a word, its meaning does not live in a single neuron or a single brain region, but in a shimmering, high-dimensional pattern distributed across the hippocampal population, a pattern whose geometry looks remarkably like the embeddings of a machine that learned to speak.

Subject of Research: Semantic encoding of spoken words by hippocampal neuron populations in humans

Article Title: A population code for semantics in human hippocampus

Article References: Franch, M., Mickiewicz, E. A., Belanger, J. L., Joiner, B., Katlowitz, K. A., Zhu, H., Chavez, A. G., Chericoni, A., Paulo, D. L., Goldman, A. M., Krishnan, V., Maheshwari, A., Bartoli, E., Kemmer, S., Piantadosi, S. T., Sheth, S. A., Provenza, N. R., Hayden, B. Y., & Hennig, J. A. (2026). A population code for semantics in human hippocampus. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02436-4

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02436-4

Keywords: hippocampus, semantics, population coding, single-neuron recordings, large language models, GPT-2, word embeddings, language comprehension, polysemy, contrastive coding, neural encoding, human neuroscience

The study, authored by Mengwen Xiao, Kangrui Du and Ziqi Lin, sits at the intersection of two research traditions that have matured rapidly over the past decade. The first is representational alignment, a family of techniques that quantify the similarity between the internal states of artificial neural networks and the activity patterns recorded from living brains. The second is chain-of-thought reasoning, the strategy by which large language models decompose complex problems into intermediate steps before committing to an answer. The new research connects these traditions in a striking way: instead of using brain activity only as a post-hoc yardstick, the authors use it as a training signal, encouraging a model’s internal representations to track those observed in human neural data while it reasons.

To appreciate why this matters, it helps to recall how the relationship between brains and machines has typically been studied. In a landmark 2023 analysis, researchers including Alex Zador and colleagues argued in Nature Communications that understanding natural intelligence requires studying both brains and their artificial analogues together, a perspective that helped galvanize the NeuroAI community. Subsequent work by Greta Tuckute and collaborators, published in Nature Human Behaviour in 2024, showed that language models vary widely in how well their internal states predict neural responses to sentences, and that this predictive power relates to how ‘brain-like’ a model’s representations are. Until recently, however, such alignment scores were largely descriptive. They told researchers which models resembled the brain, but they did not tell the models how to become more brain-like, nor whether doing so would improve them.

The conceptual shift from passive alignment to active neural guidance is the heart of the new work. In the passive paradigm, a trained model is frozen and its hidden activations are compared, layer by layer, with recordings such as functional MRI or magnetoencephalography taken while people read or solve problems. In the active paradigm described by Xiao and colleagues, the comparison happens during learning. The model receives an additional objective that rewards representational similarity to human neural data, effectively telling the network, at each stage of processing, that its internal geometry should not drift too far from the geometry observed in the brain. This turns the brain from a mirror into a compass, and the consequences for reasoning performance are substantial.

The technical machinery behind this idea draws on a rich literature. Representational similarity analysis and related encoding models provide differentiable measures of alignment, meaning that the gap between a model’s internal states and recorded brain activity can be computed as a quantity that gradient descent can minimize. Changde Du and Huiguang He, writing in an accompanying News and Views article in Nature Machine Intelligence, place the new study in the context of their own earlier work, published in the same journal in 2025, which explored how neural signals can inform machine learning systems. The new contribution extends this program from perception and language understanding into the domain of multi-step reasoning, where errors compound across steps and small representational deviations can cascade into dramatically wrong conclusions.

Reasoning is precisely where large language models remain most fragile. Chain-of-thought prompting, introduced by Jason Wei and colleagues at NeurIPS in 2022, showed that asking models to generate intermediate reasoning steps can unlock dramatic gains on arithmetic, logic and symbolic tasks. Yet subsequent analyses, including a large-scale study by Abulhair Saparov and colleagues in 2023, revealed that models often succeed on familiar reasoning patterns while failing catastrophically on novel ones, suggesting that their apparent reasoning may rest on shallow pattern matching rather than robust inference. If human neural representations encode something about the structure of valid inference that current training objectives fail to capture, then aligning with those representations during training could inject exactly the inductive bias that chain-of-thought systems lack.

The evidence reviewed by Du and He suggests that this is more than speculation. When the neural-guided objective is incorporated into training, the resulting models not only show stronger correspondence with human brain activity but also demonstrate improved reliability on reasoning benchmarks, producing more consistent and accurate chains of intermediate steps. The authors of the News and Views piece frame this as a move ‘from passive alignment to active neural guidance,’ a phrase that captures the dual payoff of the approach. Alignment ceases to be a descriptive statistic and becomes a functional ingredient of learning, and the improvement in reasoning suggests that the brain’s representational structure carries information that is genuinely useful for artificial inference, not merely an incidental byproduct of shared training data.

The finding also speaks to a long-running debate about what language models actually learn. Ev Fedorenko, Steven Piantadosi and Edward Gibson argued in Nature in 2024 that formal linguistic competence, the ability to produce fluent grammatical language, can be dissociated from functional linguistic competence, the ability to use language for thinking. Large models have largely conquered the former while remaining uneven at the latter. Neural guidance offers a possible bridge: if the internal representations that support human functional competence can be transferred, even partially, into artificial systems, then machines might acquire not just the form of reasoning but something closer to its substance. Related work by Juan Prado, Annapurni Chadha and James Booth, published in the Journal of Cognitive Neuroscience in 2011, mapped the neural substrates of arithmetic processing in children, illustrating the kind of structured neural data that such alignment approaches can exploit.

Independent work reinforces the momentum behind this direction. A 2026 study by Jie Chen, Yue Qi, Yu Wang and Guang Pan in Nature Communications reported further evidence that neural signals can inform and improve machine learning systems, indicating that the strategy is not an isolated result but part of an emerging research program. Together, these studies suggest that the coming years will see increasingly tight coupling between neuroimaging pipelines and model training loops, with brain data flowing not only into evaluation suites but into the loss functions themselves. The practical challenges remain considerable: collecting high-quality neural recordings at scale is expensive, individual differences in brain organization complicate generalization, and the causal question of why neural alignment helps reasoning is only beginning to be answered.

Still, the implications are hard to overstate. If the human brain can serve as a reliable guide for training artificial reasoners, then neuroscience and machine learning are no longer parallel enterprises exchanging occasional insights but a single, integrated discipline in which each field sharpens the other. Models guided by neural signals could become more trustworthy partners in science, medicine and education, while the act of aligning machines to minds may itself reveal which aspects of neural computation are essential to intelligence and which are biological accident. The work of Xiao, Du and Lin, and the perspective offered by Du and He, mark an early but compelling step in that direction, transforming the humble brain scan from a passive record of thought into an active instrument for teaching machines how to think.

Subject of Research: Using human brain signals to actively guide and improve the reasoning of artificial intelligence models

Article Title: Steering machine reasoning with brain signals

Article References: Du, C., & He, H. (2026). Steering machine reasoning with brain signals. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01302-z

Image Credits: AI Generated

DOI: 10.1038/s42256-026-01302-z

Keywords: NeuroAI, representational alignment, large language models, chain-of-thought reasoning, brain signals, machine learning, neural encoding, Nature Machine Intelligence, cognitive neuroscience, artificial reasoning, fMRI, brain-inspired AI