A number simplex in the human medial temporal lobe

TL;DR: Recording from populations of neurons in the human medial temporal lobe while people counted dots and did arithmetic, the authors find that numbers are not encoded along a one-dimensional "mental number line" but as vertices of a high-dimensional simplex — a geometry that maximizes flexibility and linear readout. Dot arrays and Arabic numerals get distinct simplicial codes that are nonetheless linked by a linear map; the internally computed result of an arithmetic problem is decodable while the person is still calculating; and — the part that should make you sit up — LLMs show similar simplicial number geometry, and linear transformations of operand codes model how the brain converts operands into results, in a way the authors liken to attention.

One important caveat up front: only the abstract of this preprint was available to me, so everything below is built from that plus the necessary background to make it legible. I'll flag where I'm reconstructing context versus reporting the paper's claims, and I can't give you effect sizes, subject counts, or figure numbers.

The old picture: numbers on a line

The dominant model of how brains represent quantity, going back decades of monkey and human work (Nieder, Dehaene, and others), is the mental number line: numerosity is encoded along a single compressed (roughly logarithmic) axis. Individual "number neurons" have bell-shaped tuning curves centered on preferred numerosities, with tuning width growing with magnitude — which neatly explains behavioral signatures like the distance effect (5 vs 6 is harder to discriminate than 2 vs 8) and the size effect (8 vs 9 is harder than 2 vs 3).

At the population level, this model predicts a specific geometry: the neural states for 1, 2, 3, 4, ... should trace out a one-dimensional manifold — a curve — through neural state space, with neighboring numbers close together and Weber-law compression at the high end.

This picture is elegant and explains discrimination behavior well. But the authors open with an observation that's easy to underappreciate: humans do a lot more with numbers than discriminate them. We add them, compare them, bind them to objects, switch between formats (three dots, the numeral "3", the word "three"), and compose them into procedures. A 1D manifold is a terrible substrate for that kind of flexibility, for a reason worth making precise.

Why a line is computationally cramped

The key concept is shattering dimensionality: out of all the ways you could partition a set of conditions into two groups, what fraction can a linear readout separate? If your downstream circuits are (approximately) linear decoders — a standard and biologically reasonable assumption — shattering dimensionality measures how many different tasks can be built on top of a fixed representation without rewiring it.

Points on a line are almost maximally bad at this. If the codes for 1–4 sit in order along a curve, a linear readout can implement "greater than 2" trivially — but it cannot implement "is the number even?" ({2,4} vs {1,3}), because no hyperplane separates interleaved points on a line. Every grouping that isn't an interval is inaccessible.

Now consider the opposite extreme: place the k number codes at the vertices of a simplex — the generalization of an equilateral triangle (2-simplex) and regular tetrahedron (3-simplex) to k-1 dimensions, where all k vertices are mutually equidistant. Simplex vertices are in general position, so every dichotomy is linearly separable: shattering dimensionality is maximal. Any grouping of numbers you might ever need — parity, "operands of the current problem", "greater than 5", arbitrary category labels — is one hyperplane away.

The trade-off is that a simplex throws away metric structure: on a pure simplex, 2 is no closer to 3 than to 9, so distance and size effects don't fall out of the geometry for free. The abstract describes the manifolds as "simplex-shaped," which in practice probably means simplex-like with some residual magnitude structure — but I can't confirm the mixture from the abstract alone. This is the first thing to check in the full text.

What they did and found

The data come from populations of medial temporal lobe (MTL) neurons recorded in humans — almost certainly epilepsy patients with depth electrodes, the standard setting for human single-unit work, though the abstract doesn't say — during two tasks: dot counting and arithmetic. Four headline results:

1. Numerosity coding is high-dimensional and simplicial. Population responses to different numbers form manifolds with high shattering dimensionality and expressibility, not the low-dimensional curve the number-line model predicts. Note the anatomical surprise embedded here: classical numerosity coding lives in parietal cortex (intraparietal sulcus) and prefrontal cortex. Finding robust, structured number codes in the MTL — hippocampus and its neighbors, the memory system — suggests numbers are represented there the way the MTL represents other concepts: as discrete, composable entities (think concept cells), which is exactly what a simplex code looks like.

2. Format-specific codes linked by a linear bridge. Dot arrays and Arabic numerals evoke distinct simplicial population codes — the neural state for "•••" is not the neural state for "3" — yet within a task the two codes share a latent structure that is linearly transferable: a linear map takes one simplex onto the other. This is a satisfying middle ground in a long-running debate about abstract versus notation-dependent number codes. The representation isn't format-invariant at the level of raw activity, but it's format-invariant up to a linear transformation, which is all a downstream linear readout needs (train on one format, apply a fixed rotation, decode the other).

3. Results are computed before they're reported. During the calculation period of arithmetic problems — after the operands appear, before the answer is given — the result is already decodable from MTL populations. So the code isn't just a passive stimulus label; it participates in computation. And decoding accuracy for the internally computed result correlates with individual mathematical capacity across subjects, tying the neural geometry to behavior. (The abstract gives no correlation values; treat the strength of this link as unknown until you see the scatter plot.)

4. Arithmetic as linear transformation of simplicial codes — and the LLM parallel. Two connected claims. First, number representations inside large language models exhibit similar simplicial geometry. Second, in the brain data, linear transformations of the simplicial operand representations model how operands are converted into the decodable result — a mechanism the authors say resembles the attention architecture of LLMs.

simplex vertices in, linear mixing, simplex vertex out — the same algebra an attention head applies to value vectors

Why is the linearity claim plausible at all? Here the simplex geometry does real work. If numbers 1..k are encoded as (near-)orthogonal simplex vertices v_1, \dots, v_k, then any function on pairs of numbers — including addition — can be implemented by linear operations on the (combined) operand code, because the operand pairs are themselves linearly separable conditions. On a 1D number line, "linear map from operands to result" can only express affine functions of magnitude; on a simplex, it can express arbitrary lookup tables. The high-dimensional geometry is what makes simple (linear, attention-like) computation sufficient. That, to me, is the conceptual heart of the paper: representation geometry and computation mechanism are two sides of the same design choice, in brains and in transformers.

The LLM connection has independent support worth knowing about: interpretability work has found that LLMs represent numbers with structured geometries (e.g., helical/circular components used for modular addition, and the "addition via rotation" results in mechanistic interpretability). The claim here — simplicial geometry in LLM number representations mirroring MTL geometry — is a different and complementary observation, but the abstract doesn't specify which models, which layers, or which probing methodology, so the depth of the parallel is exactly the thing to scrutinize in the full text.

How the codes fit together

A useful way to hold the whole result set in your head:

distinct population codes per format, same simplicial skeleton underneath

How strong is the evidence, and what to be skeptical about

Since I only have the abstract, I can't audit statistics, subject counts, unit yields, or controls. But here's the skeptic's checklist I'd bring to the full text:

Is "simplex" a discovery or a near-tautology? Any set of k reliably discriminable, roughly balanced categories will look approximately simplicial in a high-dimensional space — that's what "decodable and not compressed" means geometrically. The substantive claim has to be comparative: the geometry is closer to a simplex than to the compressed 1D curve that decades of number-neuron tuning data predict, with cross-validated dimensionality and shattering analyses. If the paper shows that magnitude explains only a small fraction of the between-number distance structure, that genuinely overturns the default model — at least for MTL. Which raises the next point.

MTL is not parietal cortex. The mental number line evidence comes largely from IPS and PFC. It's entirely possible that both pictures are right in different places: a compressed magnitude code in parietal cortex for perception and comparison, and a discrete simplicial "concept" code in MTL for memory, binding, and symbolic manipulation. The abstract's framing ("a high dimensional representational foundation for numerical cognition in the brain") is stronger than what MTL-only recordings can establish. Also remember the standing caveat for all human single-unit work: epilepsy patients, electrodes placed for clinical reasons, limited task time.

How deep is the LLM analogy? Two systems having simplex-like category codes is weak evidence of shared computation — one-hot-ish codes for discrete tokens are almost forced by the setting. The interesting claim is the mechanistic one: that linear transformations of operand codes model the brain's operand→result conversion, paralleling how attention linearly mixes value vectors. Whether that's demonstrated (e.g., a fitted linear model on operand-period activity predicting result-period geometry, cross-validated across problems) or merely gestured at is the difference between a genuinely novel brain–AI bridge and a decorative comparison. The abstract's own hedging — "some resemblance" — is honest and worth respecting.

Behavioral link. The correlation between result-decoding accuracy and individual math ability is potentially the most human-relevant result (imagine a neural marker of arithmetic fluency), but across-subject correlations in small clinical cohorts are notoriously fragile. Check the n.

Range of numbers. Dot-counting tasks typically use small numerosities. A simplex over 1–5 is a much weaker claim about "numerical cognition" than one spanning a decade of magnitudes, where the line-vs-simplex distinction really bites (Weber compression is what the line model predicts at larger n).

Why this matters if it holds

The mental number line has been one of cognitive neuroscience's most stable exports for thirty years. This paper says that, at least in the human MTL, the population geometry is closer to the opposite extreme: numbers as maximally separated discrete concepts, with magnitude relations and format identity carried as linear structure on top of the simplex rather than as the manifold's backbone. That reframing does real explanatory work: it explains how a fixed number code can support arbitrary novel groupings and lookups (high shattering dimensionality), how format-invariance can coexist with format-specific activity (linear transferability), and why arithmetic could be implemented by simple linear machinery (linear maps on simplex vertices can realize arbitrary tables). And it makes the brain–transformer correspondence unusually concrete: same representational geometry, same class of linear-mixing computation — the kind of convergence you'd expect if both systems face the same problem of doing flexible symbolic work with distributed vectors.

When the full text is available, go straight to the geometry analyses (shattering dimensionality and the simplex-vs-line model comparison), then the operand→result linear-transformation modeling — those two carry the paper's novelty. The LLM comparison is worth reading last, with the "geometric resemblance vs shared mechanism" question firmly in mind.