Lorenzo Posani, Shuqi Wang, Samuel P. Muscinelli, Liam Paninski and Stefano Fusi (2026), Nature, published online 15 July 2026 — DOI-10.1038/s41586-026-10668-4
Key takeaways
- Within a region, neurons rarely form types. Each selective neuron’s response is summarised by eight time-summed coefficients of a reduced-rank regression on block prior, stimulus side, contrast, choice, outcome, wheel velocity, whisking and licks. A region counts as categorical when the best k-means partition of those vectors has a higher silhouette score than 100 draws from a Gaussian with the data’s own mean and covariance (Bonferroni P < 0.05). Only VISp, AUDp and SSp-ul pass; SSp-ll and GU are near the threshold (Fig. 3d). What most regions show instead is an elongated, unclustered cloud: some variables are encoded much more strongly than others (Discussion).
- Types appear at larger scales, and follow anatomy. Pooling the non-categorical regions by anatomical module gives clusters in the somatomotor (z = 8.56), medial (7.84) and lateral (2.32) modules but not the prefrontal (0.54); the whole cortex gives z = 8.0. Cluster labels align with area labels (Fig. 3e,f). A neuron’s region can be decoded from its selectivity profile alone (0.233 against a 0.033 null), and regions with more anatomical connectivity have more similar average profiles (Spearman ρ = 0.40) (Fig. 2c–f).
- Diverse selectivity buys separable geometry. Rows and columns of one neurons × conditions matrix share a spectrum, so k tight clusters of neurons cap the participation ratio of the condition geometry near min(k, M) (Methods, Extended Data Fig. 9). Across regions, a measure of response diversity (α-diversity) tracks representation dimensionality and the fraction of balanced dichotomies of 16 conditions that a cross-validated linear readout decodes (Figs 5c, 6c).
- Over what each area does tell apart, nearly everything is readable. An iterative procedure merges conditions until every pair is decodable at ≥ 0.666, leaving M_IC from 5 (SSp-n) to 16 (MOs), increasing up the anatomical hierarchy (ρ = 0.77). Over those conditions, separability is ≥ 0.95 in 15 of the 16 regions with enough trials, GU being 0.82, and no longer depends on diversity (Fig. 6d, Extended Data Fig. 11e).
- A warning the authors draw from it. When conditions are pairwise separable, every dichotomy of them is decodable above chance, so finding one variable decodable “often lacks significance, as all other variables are also likely to be decodable” (Discussion).
Standing in the record
Second reading under ADR-013. The Anthology of the SOTA holds this
paper as ANTH-LIT-696, read in full on 2026-09-25 and tagged
concept-geometry. Its question is what the paper implies for
representations in machine learning: that a successful linear probe is
weak evidence a concept is specifically represented, and that a
Gaussian-null clustering test could ask whether a network’s units or a
sparse dictionary’s latents form functional types. This record reads it
for a different question, about brains: do neurons within a cortical
area come in discrete functional types or form a continuum of mixed
selectivity, and what does that structure do to how many distinctions the
area makes available to a linear readout? That is the empirical ground
under the mixed-selectivity account of flexible cognition (Rigotti et al.
2013; Fusi, Miller & Rigotti 2016), which this paper extends from monkey
prefrontal cortex to a brain-wide survey, and which this record does not
yet hold.
The paper does not answer a question about abstraction, though it is easy to read it that way. It measures separability, the readout’s capacity to split conditions in many arbitrary ways, and not the cross-condition generalisation by which Bernardi et al. (2020) measure abstract, low-dimensional structure. The authors say so: high separability “does not imply a complete absence of structure”, and a representation can have the generalisation properties of a disentangled one and still be maximally separable (Discussion). It is evidence about how diverse cortical codes are, not about whether they abstract.
Boundary. The paper carries no instruction for machine-learning
practice; its only contact with artificial networks is the Discussion’s
citation of RNN work on when modular structure emerges. Neuroscience is
not an anthology topic, and the question this entry asks of the paper is
one no anthology topic holds. It therefore does not carry
anthology-candidate: that flag asks whether a work should move, and this
one is held there already for the anthology’s own question, as LIT-267 and
LIT-460 are.
Connections to the record’s holdings.
- Neural collapse. Papyan, Han and Donoho (LIT-618) showed that classifiers trained past zero error collapse each class’s last-layer features onto its mean and spread the means into a simplex equiangular tight frame, which the record states as THEORY-083. That looks like the opposite of “rarely categorical”, and a referee proposed it as related work on categorical representations. It is a different sense of the word. Neural collapse is about conditions in neural space: samples of a class collapse to one point and the class means sit as far apart as possible, which in this paper’s terms is maximal M_IC with maximal dimension, the tetrahedron of its Fig. 6a that is separable in every way. This paper’s “categorical” is about neurons in conditions space, forming types. Neural collapse is invariant to rotating the feature basis and says nothing about whether units form types; by the paper’s own row/column argument, a collapsed C-class layer whose units formed k tight types would need k ≥ C − 1. The two are not in conflict, and the published reference list does not cite neural collapse.
- Decoding as evidence. Vishne et al. (LIT-644) decode image category from human ventral temporal cortex, and THEORY-102 rests on that decoding. This paper’s warning applies to how such results are read: above-chance decoding of a chosen variable is weak evidence of representing that variable when the conditions are pairwise distinguishable. Vishne et al.’s sustained near-ceiling accuracy and their exemplar-geometry analysis are stronger than above-chance decoding, so the warning qualifies the inference without undercutting it. Nothing is declared; the two papers measure different species, signals and variables.
- Readout and basis. THEORY-008 holds that what a regularised linear readout can decode depends only on a representation’s kernel, which is unchanged by rotating the units. That is why separability cannot tell “explicit” modularity (segregated neuron types) from “implicit” modularity (the same geometry rotated), as the paper’s Discussion says, and why the categoricality question needs a separate test in the neuron basis.
- Verzhbinsky et al. (LIT-633) and Xu et al. (LIT-638), filed in the recent neural-dynamics batch, study coordination across regions in humans (co-firing during ripples, spiral waves in fMRI). They share this paper’s interest in large-scale organisation but test nothing it claims, and no relation is declared.
The record holds none of the works this paper builds on or answers:
Rigotti et al. (2013), Fusi, Miller and Rigotti (2016), Bernardi et al.
(2020), Raposo et al. (2014), and Hirokawa et al. (2019), whose
orbitofrontal “categorical” result one referee calls corrected by this
paper. So no extends, corrects or compared_against is declared.