Neuroscientists investigating how different brain regions coordinate perception and action have discovered that the structural formatting of information within neuronal populations, rather than the mere presence of encoded signals, directly predicts how an area influences behavior. For decades, electrophysiological recordings across cortical circuits have revealed that sensory, motor, and cognitive variables are represented redundantly across numerous anatomical areas, raising a fundamental question regarding why the mammalian brain maintains distinct, specialized structures if broad populations encode essentially everything. In a rigorous investigation published in Nature Neuroscience, a research team led by Douglas A. Ruff, Sol K. Markman, Jason Z. Kim, and Marlene R. Cohen addressed this long-standing puzzle by evaluating how population geometry dictates causal function. By comparing the middle temporal area, a visual region known to support visual motion perception, with the dorsolateral prefrontal cortex, a frontal region implicated in reward processing, executive control, and saccadic planning, the authors demonstrated that differing population geometries determine how circuits causally transform sensory evidence into decisions.
To establish an empirical framework capable of linking multidimensional neural activity directly to continuous motor actions, the investigators trained two adult male rhesus macaques, Macaca mulatta, weighing eight and ten kilograms, to perform a continuous motion estimation task. Unlike conventional binary two-alternative forced-choice paradigms that restrict responses to discrete alternatives, this continuous design allowed the researchers to extract fine-grained behavioral distributions while requiring the subjects to simultaneously evaluate sensory motion and expected rewards. The monkeys fixated on a central visual point before viewing a circular target ring displayed on a high-refresh monitor. Dynamic random dot kinematograms appeared within a stationary aperture inside the ring at controlled motion coherences. The animals reported their estimates of motion direction by executing precise saccadic eye movements to coordinates along the target ring between 150 and 1500 milliseconds after visual stimulus onset. Juice rewards were scaled according to directional estimation accuracy and further modulated by colored sections located along the target perimeter.
The experimental architecture incorporated dynamic reward landscapes to compel the primates to integrate perceptual inputs with valuation metrics during every trial. Portions of the peripheral target ring were colored pink or purple to designate regions of increased, decreased, or neutral reward scaling, with the high-reward and low-reward centers positioned exactly 180 degrees opposite one another. The spatial locations of these designated reward centers varied systematically across experimental sessions according to the directional tuning profiles of the recorded neural populations. Analysis of the primates’ eye movements confirmed that their behavioral choices varied continuously with stimulus motion while exhibiting consistent systematic biases toward the spatial center associated with higher juice volume. Furthermore, trials where choices were directed toward regions of elevated reward scaling exhibited systematically faster saccadic reaction times. This behavioral pattern confirmed that the subjects were actively combining noisy incoming sensory motion evidence with cognitive reward expectations into unified motor plans rather than following rigid perceptual rules.
While the animals executed this continuous integration task, the researchers performed simultaneous multi-channel electrophysiological recordings across dozens of isolated single units in both the middle temporal visual area and the dorsolateral prefrontal cortex. Using linear multi-contact probes advanced into the cortex via stereotaxically positioned chambers, the team captured extracellular spiking activity, local field potentials, and task events at thirty thousand samples per second. Both recorded populations exhibited robust modulations driven by visual motion direction as well as the prevailing reward condition, confirming that both sensory and value variables penetrated visual sensory and frontal executive regions. However, dimensionality reduction and demixed principal component analysis exposed profound differences in how these signals were geometrically formatted within neural state space. In the middle temporal area, the axes accounting for visual motion direction and reward condition were arranged orthogonally, maintaining separable, independent dimensions. In sharp contrast, dorsolateral prefrontal cortical populations organized the variables jointly, exhibiting pronounced nonlinear mixed selectivity that collapsed motion and reward into shared, decision-aligned population representations.
Quantitative statistical evaluations confirmed that the angular separation between the primary motion and reward dimensions in neural population space was significantly larger in the middle temporal area than in the dorsolateral prefrontal cortex across fifteen simultaneous recording sessions, as verified by Wilcoxon’s signed-rank tests. Furthermore, application of demixed principal component analysis demonstrated that a significantly greater proportion of total population variance was explained by the mathematical interaction between motion direction and reward condition in the prefrontal cortex than in the visual middle temporal cortex. To formulate explicit causal hypotheses regarding how these divergent representational geometries influence downstream behavioral choices, the team constructed a two-stage recurrent neural network model. The artificial architecture featured an upstream visual module trained to sharpen noisy motion signals and output motion direction and reward centers along independent, separable output dimensions. The recurrent weights of this visual module were then frozen, and its outputs were routed into a downstream prefrontal module trained to calculate expected reward values across all potential directional choices.
The artificial network architecture operated under strict biological constraints, including Dale’s law governing positive excitatory and negative inhibitory synaptic projections, with recurrent hidden units approaching stable dynamical fixed points at trial conclusion. Crucially, the visual module in the artificial network accounted for 98.0 percent of directional variance within its first two principal components and 55.3 percent of reward variance along its primary reward axis without cross-variable interaction variance. Conversely, the prefrontal module captured task variables in a higher-dimensional format, with twenty percent of explained variance across task conditions driven specifically by nonlinear motion-reward interactions. When the investigators simulated targeted perturbations by injecting synthetic depolarizing input into individual recurrent units, the two modules exhibited fundamentally divergent behavioral consequences. Perturbing units within the middle temporal module systematically biased choices toward the preferred direction of the activated unit via vector averaging, provided angular disparities remained within 150 degrees. Conversely, stimulating the prefrontal module produced bimodal, winner-take-all choices without vector averaging, selecting either the visual motion direction or the unit’s preference.
Armed with these computational predictions, the researchers conducted causal microstimulation experiments in the behaving monkeys by delivering biphasic electrical currents at 200 Hertz across selected probe contacts on randomly interleaved trials. In full agreement with the neural network predictions, causal electrical microstimulation delivered to the middle temporal area induced classic vector averaging, systematically deflecting saccadic choices toward the receptive field preference of the stimulated electrode site across fifty stimulated sites from twenty-nine experimental sessions. Strikingly, the magnitude of this directional bias remained statistically indistinguishable regardless of whether the preference of the stimulated visual site aligned with the increased or decreased reward sector, yielding a two-tailed Wilcoxon rank-sum test probability value of 0.64. Because the middle temporal cortex formats motion and reward information along separable population axes, artificially elevating local firing rates modified directional estimates without interacting with cognitive reward expectation, leaving behavioral weighting across reward conditions entirely invariant.
Microstimulating the dorsolateral prefrontal cortex across thirty-eight sites during twenty-three sessions generated an entirely distinct behavioral phenotype that matched the winner-take-all dynamics forecasted by the recurrent model. Instead of producing intermediate vector-averaged choices, prefrontal microstimulation caused the monkeys to alternate abruptly between reporting the veridical visual stimulus direction and selecting the response field preference of the stimulated site. Crucially, this prefrontal causal intervention exhibited a pronounced, statistically significant interaction with the prevailing reward condition, producing substantially larger behavioral shifts when the preference of the stimulated site aligned with the region of increased reward scaling compared to the decreased reward condition, with Wilcoxon rank-sum testing yielding a significance value below 0.01. Because the prefrontal cortex formats motion and reward jointly within an integrated decision space, external electrical perturbation could not act upon motion representation independently of reward expectation, forcing choices into discrete, value-dependent attractors that governed the ultimate motor execution.
To eliminate potential confounding variables and verify that these divergent behavioral outcomes did not stem from trivial methodological disparities, the authors executed rigorous control experiments. First, they addressed whether the observed differences arose from disparities in the relative potency of electrical microstimulation across cortical regions by titrating visual stimulus strength through lower motion coherences. Decreasing motion coherence amplified the magnitude of behavioral effects in both brain areas without altering the qualitative nature of their contributions: middle temporal stimulation continued to induce smooth vector averaging, whereas prefrontal stimulation preserved its bimodal winner-take-all structure. Second, the authors tested whether spatial receptive field alignments accounted for the results, as visual stimuli typically fell within middle temporal receptive fields while target rings overlapped prefrontal response fields. Stimulating middle temporal sites whose receptive fields overlapped the peripheral target ring rather than the visual motion stimulus failed to elicit systematic directional biases or winner-take-all selections, confirming that the effects reflected intrinsic information formatting.
Additional regression analyses removing choice-related activity demonstrated that the fundamental contrast between separable and joint population representations persisted independently of motor execution signals, proving that prefrontal population geometry does not merely reflect finalized action commands. These collective findings provide direct empirical and computational evidence resolving why redundant informational codes permeate mammalian cortical structures. Although multiple brain regions simultaneously register identical environmental inputs, their distinct representational geometries dictate whether they provide veridical, uncorrupted sensory estimates or flexibly reformat variables into decision-oriented configurations. By shifting the scientific focus from basic neural decoding to the geometric formatting of population activity, the authors establish an experimental paradigm for predicting how less understood brain areas causally shape complex behavior. This mechanistic framework demonstrates that modularity in cortical architecture endures not because regions monopolize specific information types, but because different functional geometries are computationally necessary to transform sensory perception into adaptive, goal-directed actions.
Subject of Research: Neuronal population information formatting and causal contributions to behavior
Article Title: Information formatting in neuronal populations predicts differential functional contributions to behavior
Article References: Ruff, D. A., Markman, S. K., Kim, J. Z., & Cohen, M. R. (2026). Information formatting in neuronal populations predicts differential functional contributions to behavior. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02514-7
Image Credits: AI Generated
DOI: 10.1038/s41593-026-02514-7
Keywords: neural population formatting, middle temporal area, dorsolateral prefrontal cortex, microstimulation, decision-making, rhesus macaques, recurrent neural networks, continuous estimation task, vector averaging, winner-take-all, mixed selectivity, reward expectation, motion perception, computational neuroscience, electrophysiology, demixed PCA
News Source: Cassandra Pierce. (November 12, 2026). Brain Regions Format Shared Information Differently to Steer Primate Behavioral Choices. Scienmag.
