Source
- Posani, Wang, Muscinelli, Paninski & Fusi (2026), LIT-708, read in full in NOTE-549: Figs 2, 3 and 6, Extended Data Figs 6, 8, 9 and 11, the Methods on clustering, independent conditions and separability, and the Peer Review File.
What was actually shown
The setting. Posani et al. (LIT-708) analysed the International Brain Laboratory’s Brainwide Map: about 14,000 Neuropixels units from 43 regions of mouse cortex, recorded while trained mice turned a wheel to bring a visual stimulus, shown on the left or right, to the centre of a screen, in blocks of trials where one side was more likely. Each neuron’s selectivity is eight coefficients from a regression of its activity on block prior, stimulus side, contrast, choice, outcome, wheel velocity, whisking and licking, summed over the trial. Only the 4,617 neurons the regression explains better than a trial average are kept.
The first half: no types within an area. An area counts as categorical when k-means finds clusters of these eight-dimensional profiles with a higher silhouette score than the same number of points drawn from a Gaussian with the data’s own mean and covariance, at Bonferroni P < 0.05. Of the about 20 areas with at least 50 selective neurons, only primary visual (VISp), primary auditory (AUDp) and the upper-limb field of primary somatosensory cortex (SSp-ul) pass. Even there the clusters are weak: VISp’s silhouette is about 0.23 against a null near 0.14–0.18 (Fig. 3b,d). Other areas show an elongated cloud, with some variables encoded much more strongly than others, and no separated groups. The finding holds across selectivity thresholds, a second clustering algorithm and time-resolved profiles, and two other tests (clustering raw firing rates over 16 conditions, and ePAIRS) also find only two categorical areas each (Extended Data Fig. 6).
Types exist, at a larger scale. Pooled across an anatomical module or the whole cortex, profiles do cluster (whole cortex z = 8.0; somatomotor 8.56, medial 7.84, lateral 2.32, prefrontal 0.54, not significant), and the clusters align with area labels (Fig. 3e,f). A single neuron’s area can be decoded from its profile, and areas with more anatomical connectivity have more similar profiles (Fig. 2d–f). Specialisation is real between areas. What is rare is discrete types within one.
The second half: separable in nearly every way. For geometry, four binarised variables (whisking, block, stimulus side, contrast) define 16 conditions. For each area, conditions are merged until every remaining pair is linearly decodable at ≥ 0.666. Between 5 (SSp-n) and 16 (MOs) remain, more in areas higher in the anatomical hierarchy (ρ = 0.77). Over those conditions, a cross-validated linear readout on a resampled 4,000-neuron population decodes ≥ 95% of 200 random balanced dichotomies above the 99th percentile of a label-shuffled null in 15 of 16 areas; the gustatory cortex reaches 0.82 (Fig. 6d, Extended Data Fig. 11e). Mean accuracy over those dichotomies runs from 0.65 to 0.90.
Why the halves go together. The neurons × conditions activity matrix has the same spectrum read by rows (neurons in conditions space) or by columns (conditions in neural space). If neurons formed k tight types, the conditions would span about k dimensions at most; the paper derives the participation ratio of Gaussian clusters, which tends to min(k, M) as the clusters tighten, and fits it to the regions (Methods, Extended Data Fig. 9). Diverse, unclustered profiles leave room for the conditions to sit in general position, which is what a readout needs to split them arbitrarily (Cover’s theorem, here in cross-validated form). The data agree: over all 16 conditions, areas with more diverse profiles have higher-dimensional geometry and higher separability (ρ = 0.67 and 0.69). Over independent conditions separability is near ceiling everywhere, and no longer depends on diversity.
What this says about cognition. It is the brain-wide version of the claim behind mixed selectivity: that cortex favours diverse, mixed responses because they let simple readouts compute many different functions of the same variables. The authors’ own conclusion is that cortex “prioritize[s] diversity over categorical structure”. That is an interpretation; what is shown is the conjunction in the title.
What this does not say
- It does not say cortex has no functional specialisation. Areas differ in what they encode and how strongly. The number of distinguishable conditions ranges threefold, and types appear when areas are pooled.
- It does not say there are no neuron types. It concerns response profiles over eight variables of one task. Tuning to anything the task does not vary is invisible (Referee 3’s point, now in the paper’s Limitations). Transcriptomic types are not examined, and the authors list continuous or only partly separated cell types as one explanation.
- “Rarely categorical” means failing a Gaussian null. The authors concede this is a narrow class: an elongated cloud with strongly selective neurons on its tails fails the null too. A neuron strongly selective for a stimulus category (Freedman’s sense) can sit in such a cloud.
- “Separable” is lenient. Above a shuffle null near 0.53 accuracy, on resampled pseudopopulations. Conditions were first merged until every pair was decodable, which removes the commonest route to failure; what remains is a test against near-degenerate geometry such as collinear conditions (Fig. 6a). The synthetic results show it is blind to anisotropy (Extended Data Fig. 8h,i).
- It does not say the geometry is high-dimensional in absolute terms. The participation ratio over independent conditions is at most 5.3, in MOs with 16 conditions.
- It says nothing about abstraction. Separability measures how many ways a readout can split the conditions, not whether a code generalises across conditions. The authors note that a disentangled, abstract code can also be maximally separable.
- The hierarchy trends are weaker than the two halves. Categoricality falls along the hierarchy (ρ = −0.62) in the main pipeline but not in the conditions-space one, and ePAIRS flags a high-hierarchy area (MOs).
Connections
- THEORY-083 (neural collapse) looks like the opposite claim and is not. Papyan, Han and Donoho (LIT-618) found that a classifier trained past zero error collapses each class’s last-layer features to its mean and spreads the means into a simplex equiangular tight frame. That is a claim about conditions in feature space, and in this account’s terms it is maximal distinguishable conditions with maximal dimension, the arrangement separable in every way. This account’s “categorical” is about neurons forming types, on which neural collapse is silent: it is unchanged by rotating the units. By the row/column bound, a collapsed C-class layer whose units formed k tight types would need k ≥ C − 1. No relation is declared: the two accounts are about different phenomena, in different systems.
- THEORY-008 holds that what a regularised linear readout can decode depends only on the representation’s kernel, which a rotation of units leaves unchanged. That is why the second half cannot speak to the first: separability is identical whether the same geometry is carried by segregated types or by mixed units (“explicit” versus “implicit” modularity in the paper’s Discussion). The two halves need two tests.
- THEORY-102 rests on decoding image category from human ventral temporal cortex (Vishne et al., LIT-644). The source paper draws a warning from the second half: when conditions are pairwise separable, nearly any grouping of them is decodable, so decoding one chosen variable above chance is weak evidence that it is specifically represented. THEORY-102’s evidence is near-ceiling, time-generalising decoding and exemplar geometry, so the warning qualifies how such evidence is read without undercutting it.