Polytope (Simplex)

structures / Manifolds / Polytope (Simplex)

OBJECTgeometric-objectsubsetmanifold-with-cornerscornersconvexadvancedhow it's classified →

The convex hull of k vector representations, one per value of a mutually-exclusive (categorical) concept — e.g. {mammal, bird, reptile, fish}. Generically a (k-1)-simplex, not just k independent directions: unlike a binary feature (one direction), a categorical feature needs a magnitude-bearing vector per value so that differences between values remain meaningful.

Replicationcomputed from the corpus — never hand-assigned

21 papers · no shared authors8 architecture classes · across papers4 domains · across papers21 model families · across papers

Filled = two or more values reported by papers that share no author — replication. Outlined = two or more values, but all from a single study — breadth, not replication. Grey = a single value. Derived from paper authorship and each model's architecture class, domain and family; it updates itself when a paper is added.

Definition

Given a categorical concept $W = \{w_0, \ldots, w_{k-1}\}$ (mutually exclusive values — a word is exactly one of them, never several at once) with a vector representation $\bar\ell_{w_i} \in \mathbb{R}^d$ for each value $w_i$ (a genuine point, with magnitude — not merely a unit direction), the polytope representation of $W$ is the convex hull

$\mathrm{Poly}(W) = \mathrm{conv}\{\bar\ell_{w_0}, \ldots, \bar\ell_{w_{k-1}}\} = \Big\{\textstyle\sum_i c_i \bar\ell_{w_i} : c_i \geq 0,\ \sum_i c_i = 1\Big\}.$

For a “natural” categorical concept (one whose values are evenly balanced against everything else in the vocabulary), this polytope is generically a $(k-1)$ -dimensional simplex — i.e. the $k$ vertices are affinely independent, the minimal-complexity polytope on $k$ points.

Intuition

A binary feature (is-animal, yes/no) needs only one direction: the projection’s sign tells you which side you’re on. But a categorical feature with $k \geq 3$ mutually exclusive values ({mammal, bird, reptile, fish}) cannot be captured by a single direction — you need $k$ actual points (not just directions) so that differences between any two of them are still meaningful vectors, exactly like the binary case’s single contrast direction $\bar\ell_{w_1} - \bar\ell_{w_0}$ . Stitching $k$ such points together by taking their convex hull gives the simplest possible shape that has all $k$ of them as extreme points: a simplex.

Properties

Requires vectors, not just directions. A direction (unit vector, no magnitude, no fixed origin) is enough for a binary contrast — a categorical concept additionally needs each value’s representation to be an actual point in $\mathbb{R}^d$ , since the polytope’s shape depends on the vertices’ positions relative to each other, not just their direction from the origin.
Vertices, not an arbitrary point cloud. Each category value is an extreme point of the polytope (not expressible as a convex combination of the others) — this is what makes “categorical concept” and “convex hull of $k$ points” the same object, rather than $k$ points scattered with no particular relationship.
Generically a simplex. With $k$ affinely independent vertices in general position, $\mathrm{Poly}(W)$ has dimension $k-1$ and is combinatorially the simplest polytope on $k$ vertices (every vertex connects to every other). Degenerate configurations (vertices not affinely independent) would collapse this to a lower-dimensional polytope instead.
Composable with hierarchy via direct sums. When categorical concepts sit at different levels of a semantic hierarchy (e.g. {mammal, bird, …} one level below animal), the contrast directions between siblings at one level are orthogonal to the contrast directions at an ancestor level — see the hierarchical-orthogonality Observation for the precise statement. The polytope itself is the per-level shape; hierarchy is what orthogonally stacks multiple such polytopes together.
Distinct from a polyhedral cone. Cone‘s polyhedral case is an unbounded wedge of non-negative combinations of basis rays (no vertices, extends to infinity); a polytope here is the bounded convex hull of a finite vertex set (extreme points are the category values themselves, not directions you scale).

Key evidence

The strongest, most specific version of this structure is the Simplex Equiangular Tight Frame (ETF): not merely a simplex, but the unique maximally symmetric one, with all vertices equidistant from the centroid (equinormal) and all pairwise angles equal (maximally equiangular). Papyan, Han & Donoho (PNAS 2020) show that standard image classifiers (VGG, ResNet, DenseNet, across seven canonical datasets from MNIST to ImageNet) reliably converge to exactly this configuration in their last-layer class-mean geometry during the terminal phase of training — a phenomenon they term Neural Collapse — and prove, via an independent information-theoretic optimal-code argument, that the Simplex ETF is the unique optimal answer to a natural question (“what class-mean geometry maximizes worst-case discriminability under a norm constraint?”), not just an empirically observed accident. See neural-collapse-last-layer-class-means-converge-to-a-simplex-equiangular-tight-frame. Caveat (Yang, Steinhardt & Hu, 2023): the collapse is real but not complete — a small, information-bearing “fine-grained structure” survives inside each apparently-collapsed class cluster (e.g. unsupervised clustering of “collapsed” CIFAR-10 representations trained on only 5 coarse super-classes recovers the original 10 classes at 93% accuracy), so each class-mean vertex should be read as the center of a genuine residual point cloud, not a literal single point.

Directional (task-relative) generalization to self-supervised encoders. Luthra, Salunkhe & Galanti (2026) extend Neural Collapse beyond a fixed supervised label set: their directional class-distance-normalized variance metric projects within-class variance onto each downstream task’s own class-separating axis before measuring collapse, rather than measuring isotropic collapse against one fixed set of classes. Across SimCLR, VICReg, MAE and DINOv2 encoders, this reveals that self-supervised pretraining collapses variance selectively along task-relevant directions while leaving orthogonal, nuisance directions largely intact — and when this holds simultaneously across many independently-sampled label tasks, the resulting per-task decision axes become nearly orthogonal to one another, geometrically explaining why one shared SSL encoder supports strong few-shot transfer across many unrelated downstream tasks at once. See directional-neural-collapse-in-ssl-encoders-explains-few-shot-transfer-via-near-orthogonal-decision-axes.

Extending Neural Collapse into causal language modeling. Wu & Papyan (2024) reframe next-token prediction as an extreme, highly imbalanced classification problem and show Neural-Collapse-style simplex-ETF signatures (NC1-NC5) emerge in the last-layer next-token representations of custom-trained GPT-2-style causal transformers, strengthening with model width, depth, and training — the first systematic demonstration that this geometry generalizes from image classifiers into generative language models. See wu-papyan-2024-neural-collapse-signatures-emerge-in-causal-language-models-scaling-with-width-depth-and-training. Du, He & Su (2026) generalize the underlying theory further, proving a “symmetry transfer” principle — permutation symmetry among prompt completions provably forces Simplex-ETF geometry, cyclic-shift symmetry instead forces circulant matrix geometry — and confirm both predictions by measuring real Gram matrices of Mistral-7B-Instruct-v0.3’s output-projection weights and context embeddings on naturally symmetric prompt sets (days of the week, months of the year). See du-etal-2026-permutation-and-cyclic-shift-prompt-symmetries-provably-transfer-into-simplex-etf-and-circulant-embedding-geometry-confirmed-in-mistral-7b.

Manifold capacity theory: from class means to whole class manifolds. Chou, Kirsanov, Yang & Chung (2026) and Kirsanov, Chou, Cho & Chung (2025) apply mean-field manifold capacity theory (Chung, Lee & Sompolinsky) to treat each category not as a single class mean but as a full “object manifold” with its own capacity, effective dimension, radius, and inter-manifold correlation structure. In CNN/ViT image classifiers, low effective manifold dimensionality and utility (“feature overspecialization”) predicts poor out-of-distribution generalization earlier and more reliably than in-distribution accuracy; in Llama-3.1-8B and Gemma-2-2B, zero-shot instructions, few-shot demonstrations, and soft prompt-tuning are shown to reorganize residual-stream category-manifold geometry through mechanistically distinct routes despite similar downstream accuracy. See low-effective-manifold-dimensionality-and-utility-predict-poor-out-of-distribution-generalization-across-cnn-and-vit-classifiers and zero-shot-few-shot-and-soft-prompt-tuning-reorganize-llm-category-manifold-geometry-through-distinct-discovered-mechanisms-despite-similar-downstream-accuracy.

Extending to control/regression settings with no explicit label set. Qi, Yin & Yang (2024) show a control-oriented analogue of Neural Collapse in image-based control (behavior cloning): with no explicit label set, they derive implicit classes from the sign pattern of a relative-pose vector (a “relative pose orthant,” REPO), and find visual representations prevalently cluster by REPO membership just as classifier representations cluster by label. Explicitly regularizing pretraining toward this clustering causally improves low-data test-time control performance by 10-35%, including in real-world planar-pushing experiments. See visual-representations-in-behavior-cloned-control-policies-cluster-by-relative-pose-orthant-and-nc-regularized-pretraining-causally-improves-low-data-control-performance.

Manifold capacity theory applied across training, reasoning, and speech. Stephenson, Padhy, Ganesh, Hui, Tang & Chung (2021) track object-manifold radius and dimension jointly across layers and training epochs of real ResNet-18/34/50 classifiers, showing these purely geometric statistics track memorization directly and reveal double descent as a genuine geometric transition. See stephenson-etal-2021-object-manifold-radius-and-dimension-track-memorization-and-double-descent-across-training-in-real-resnets. Chun, Polo & Chung (2026) apply the same framework during Chain-of-Thought reasoning in real LLMs (Ministral 3 8B Reasoning, Qwen2.5-14B-Instruct, gpt-oss-20b, Qwen2.5-7B-Instruct), finding concept manifolds transiently untangle into linearly-separable configurations right before a reasoning step then recompress afterward — a dynamic pattern invisible to static linear-probe accuracy. See chun-etal-2026-concept-manifolds-transiently-untangle-into-linearly-separable-subspaces-right-before-a-reasoning-step-then-recompress-diverging-from-probe-accuracy. Stephenson, Feather, Padhy, Elibol, Tang, McDermott & Chung (2019) apply the same framework in the speech modality, showing phoneme, speaker, and word manifolds in a real trained DeepSpeech2 network progressively untangle across layers and processing time steps. See stephenson-etal-2019-manifold-capacity-analysis-shows-deepspeech2-untangles-phoneme-speaker-and-word-manifolds-progressively-across-layers-and-time.

Foundational precedent for polytopes beyond the simplex. Elhage, Hume, Olsson, Schiefer, Henighan, Kravec, Hatfield-Dodds, Lasenby, Drain, Chen, Grosse, McCandlish, Kaplan, Amodei, Wattenberg & Olah (2022, Anthropic) — “Toy Models of Superposition,” the founding paper of the superposition/dictionary-learning research program this map’s SAE methods descend from — show that sparse features compressed into a shared dimension (“superposition”) in a toy ReLU autoencoder organize into small uniform polytopes beyond the simplex: antipodal pairs, pentagons, tetrahedra, digons, and square antiprisms, at quantized “sticky” values of a per-feature fractional-dimensionality metric $D_i = \|W_i\|^2 \big/ \sum_j (\hat W_i \cdot W_j)^2$ (e.g. $3/4$ tetrahedron, $2/3$ triangle, $1/2$ antipodal pair, $2/5$ pentagon, $3/8$ square antiprism). Causally, adversarial-example vulnerability rises over 3x as superposition forms and tracks the reciprocal of this dimensionality, and adversarial training causally reduces (though does not cleanly eliminate) the effect. See superposed-features-in-a-toy-relu-autoencoder-arrange-into-uniform-polytopes-with-quantized-fractional-dimensionality-and-vulnerability-to-adversarial-examples-tracks-that-dimensionality.

Cross-modal application: the simplex as an alignment measure, not a categorical-value structure. Cicchetti, Grassucci & Comminiello (2025) apply this same 2-simplex structure in a different role: rather than representing $k$ mutually-exclusive values of one categorical concept, the three vertices are one sample’s three different-modality embeddings (video, audio, text), and the simplex’s area — computed via the same Gram-determinant formula used generically for any simplex’s volume — serves as an anchor-free, fusion-free three-way alignment score, minimized directly as a contrastive training objective. See Triangle-area tri-modal similarity and triangle-area-similarity-jointly-aligns-three-modalities-and-beats-cosine-and-volume-based-losses-on-retrieval.

Hard-negative sampling as a causal lever on Neural- vs. Dimensional-Collapse. Jiang, Nguyen, Aeron & Ishwar (2024) train a real ResNet-50 from scratch (400 epochs, Adam) on real CIFAR-100 (and CIFAR-10, TinyImageNet) under hard-negative supervised/unsupervised contrastive losses, finding unit-ball/ unit-sphere feature normalization combined with hard-negative sampling reliably drives the trained representations to the Neural-Collapse Simplex-ETF geometry (measured via the zero-sum, unit-norm, and equal-inner-product metrics), consistently across batch sizes from 64 to 512 — while omitting either ingredient instead yields Dimensional-Collapse (see Dimensional collapse), a qualitatively different failure mode of the same training setup. See jiang-etal-2023-a-real-resnet-50-trained-with-hard-negative-contrastive-losses-and-unit-ball-normalization-converges-to-neural-collapse-etf-geometry-on-cifar100-while-omitting-either-yields-dimensional-collapse-instead.

Multi-label generalization: prototypes obey a frequency-weighted synthesis rule, not uniform averaging. Ma, Yang, Li & Hui (2026) extend Neural Collapse to real trained networks under multi-label, class-imbalanced supervision: a real ResNet-18 trained on a multi-label CIFAR-10 construction has its higher-multiplicity class prototypes obey a class-frequency-weighted synthesis rule (confirmed via a Gram-residual alignment test: residual 36.14 to 0.116 under the correctly-directed $D^{-1/2}$ rescaling versus $2.33\times10^5$ under the wrong direction) rather than the naive uniform average of single-label means, with the classical NC2/NC3 collapse metrics moving systematically as label imbalance is varied. See ma-etal-2026-a-real-resnet-18-trained-on-multi-label-cifar10-obeys-a-class-frequency-weighted-prototype-synthesis-rule-rather-than-uniform-averaging-with-collapse-metrics-tracking-label-imbalance.

Exercises

Base

Why does a binary feature (is-animal, yes/no) only need a single direction, while a 4-way categorical feature ({mammal, bird, reptile, fish}) needs 4 separate vector representations rather than, say, 4 directions from the origin?

Solution

For a binary feature, the only thing that matters is the contrast between the two values — a single direction (the difference vector) suffices to ask “which side of this hyperplane.” For 4 mutually exclusive values, you need every pairwise contrast (mammal vs. bird, mammal vs. reptile, bird vs. fish, …) to be meaningful simultaneously, and that requires each value to be a genuine point (with a fixed position, not just a direction from an arbitrary origin) so that all $\binom{4}{2}=6$ pairwise differences are well-defined vectors consistent with one shared geometric configuration — a single shared direction per value cannot encode 4 mutually consistent pairwise contrasts on its own.

A categorical concept has 3 values whose vector representations are NOT affinely independent (one is exactly the midpoint of the other two). What is $\dim(\mathrm{Poly}(W))$ in this case, and is it still a 2-simplex (triangle)?

Solution

If one vertex is the affine combination (midpoint) of the other two, all three points are collinear, so the convex hull is a line segment, not a triangle: $\dim(\mathrm{Poly}(W)) = 1$ , not 2. This is the degenerate case the “generically a simplex” claim excludes — genuinely categorical (non-collapsed) structure requires affine independence of the vertices.

Middle

Prove that every vertex of $\mathrm{Poly}(W) = \mathrm{conv}\{\bar\ell_{w_0}, \ldots, \bar\ell_{w_{k-1}}\}$ is one of the $\bar\ell_{w_i}$ themselves (i.e. no $\bar\ell_{w_i}$ is redundant), given that the $\bar\ell_{w_i}$ are affinely independent.

Solution

Suppose for contradiction some $\bar\ell_{w_j}$ is not an extreme point of the convex hull — then $\bar\ell_{w_j} = \sum_{i\neq j} c_i \bar\ell_{w_i}$ for some $c_i \geq 0$ with $\sum_i c_i = 1$ (it’s expressible as a convex, hence affine, combination of the others). But this directly contradicts affine independence of $\{\bar\ell_{w_0},\ldots,\bar\ell_{w_{k-1}}\}$ , which requires that no point in the set is an affine combination of the others. Hence every $\bar\ell_{w_i}$ must be an extreme point (vertex) of the hull.

Pro

Two categorical concepts $W = \{w_0,w_1,w_2\}$ and $Z=\{z_0,z_1\}$ sit at parent and child levels of a hierarchy (every $z$ is a $w_1$ , say). Given Hierarchical Orthogonality (every contrast direction within $Z$ is orthogonal to every contrast direction within $W$ ), what does this imply about the subspace spanned by $\mathrm{Poly}(W)$ ‘s edges versus $\mathrm{Poly}(Z)$ ‘s edge, and why is this a direct sum rather than just “two unrelated polytopes somewhere in $\mathbb{R}^d$ ”?

Solution

$\mathrm{Poly}(W)$ ’s edges are spanned by contrast directions among $\{w_0,w_1,w_2\}$ (a 2-dimensional subspace, generically); $\mathrm{Poly}(Z)$ ‘s single edge is the one contrast direction $\bar\ell_{z_1}-\bar\ell_{z_0}$ . Orthogonality of every such $W$ -contrast to the $Z$ -contrast means these two subspaces intersect only at the origin and every vector in one is perpendicular to every vector in the other — the defining property of an internal direct sum of subspaces ( $U \oplus V$ , not just $U + V$ , which would only require them to together span a bigger space without the perpendicularity or trivial-intersection guarantee). This is what licenses treating the parent-level and child-level categorical structure as independently manipulable pieces of one shared representation, rather than entangled distortions of each other.

Found in (19 observations · 20 families)

Gemma

Finding Belief Geometries with Sparse Autoencoders (2026)measured

SAE latents in Gemma-2-9B show preliminary simplex-structured belief geometry

Details

Levinson extends the belief-state-simplex framework from toy transformers to a real LLM's SAE latent space, fitting 13 priority latent clusters in a Gemma-2-9B layer-20 JumpReLU SAE with archetypal simplex fitting against 3 null-cluster controls [levinson-2026] On the barycentric predictive test, 5 of 13 real clusters show a significant advantage over the best single latent versus 0 of 3 null clusters, indicating recovered structure rather than a tiling artifact [levinson-2026] Causal steering along vertex directions gives positive scores for all 8 qualifying real clusters, but null-cluster scores overlap substantially, limiting discriminative power [levinson-2026] Only cluster 768_596 shows joint predictive and causal convergence, and the author frames the evidence as preliminary (single model and layer, modest effect sizes, a phantom vertex) [levinson-2026]

models: Gemma-2-9B · method: Sparse Autoencoders (SAE), k-Subspace Clustering, Archetypal Analysis / Simplex Fitting (AANet), Barycentric Predictive Advantage Test, Causal interventions (steering)

The Geometry of Categorical and Hierarchical Concepts in Large Language Models (2024), When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs (2026)measured

Categorical concepts are represented as simplex polytopes

Details

Park et al. show a mutually-exclusive categorical concept (e.g. mammal/bird/reptile/fish) is represented as the convex hull of its values' vector representations, generically a simplex, proven formally and confirmed in Gemma-2B and Llama-3-8B on 900+ WordNet concepts [park-etal-2024] Adding a concept's vector representation to context shifts logits for that concept's members while leaving unrelated concepts unchanged, a causal validation [park-etal-2024] Marinov et al. independently confirm the same simplex shape for language identity via a causal inner product across Qwen3-4B, Mistral-7B-v0.3, and Llama-3-8B, with four same-family languages at the vertices of a 3D simplex [marinov-etal-2026]

models: Gemma-2B · method: Linear probing, Activation Steering (Addition), Geometric analysis, Causal inner product

The Geometry of Categorical and Hierarchical Concepts in Large Language Models (2024), When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs (2026)measured

Semantic hierarchy is encoded as orthogonality between concept levels

Details

Park et al. prove (Theorem 8) that sibling-contrast directions at one hierarchy level are orthogonal to the parent's contrast direction and to any ancestor level [park-etal-2024] Combined with per-level categorical polytopes, this gives a direct-sum-of-polytopes structure where each level occupies its own orthogonal subspace, confirmed in Gemma-2B and Llama-3-8B (cosine near zero versus a shuffled-labels control) [park-etal-2024] The authors argue ordinary high-dimensional Euclidean space suffices, explicitly contrasting with hyperbolic-embedding accounts of hierarchy [park-etal-2024] Marinov et al. confirm the same cross-level orthogonality for language family versus individual language across Qwen3-4B, Mistral-7B-v0.3, and Llama-3-8B via a causal inner product [marinov-etal-2026]

models: Gemma-2B · method: Geometric analysis, Causal inner product

The Geometry of Prompting: Unveiling Distinct Mechanisms of Task Adaptation in Language Models (2025)measured

Prompting methods reorganize category-manifold geometry through distinct mechanisms

Details

Kirsanov et al. apply mean-field manifold capacity theory to residual-stream activations of Llama-3.1-8B and Gemma-2-2B under zero-shot, few-shot, and soft-prompt-tuning adaptation [kirsanov-etal-2025-geometry-of-prompting] Despite similar downstream accuracy, the three methods evolve manifold geometry through mechanistically distinct routes: some shrink effective dimension, others change inter-manifold correlation, with effects concentrated in different layer ranges [kirsanov-etal-2025-geometry-of-prompting] This decouples task performance from internal representational organization, showing behaviorally-equivalent adaptation methods rest on geometrically distinct solutions; observational-diagnostic [kirsanov-etal-2025-geometry-of-prompting]

models: Gemma-2-2B · method: Neural manifold capacity analysis

Llama

The Geometry of Categorical and Hierarchical Concepts in Large Language Models (2024), When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs (2026)measured

Categorical concepts are represented as simplex polytopes

Details

models: Llama-3-8B · method: Linear probing, Activation Steering (Addition), Geometric analysis, Causal inner product

The Geometry of Categorical and Hierarchical Concepts in Large Language Models (2024), When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs (2026)measured

Semantic hierarchy is encoded as orthogonality between concept levels

Details

models: Llama-3-8B · method: Geometric analysis, Causal inner product

The Geometry of Prompting: Unveiling Distinct Mechanisms of Task Adaptation in Language Models (2025)measured

Prompting methods reorganize category-manifold geometry through distinct mechanisms

Details

models: Llama-3.1-8B · method: Neural manifold capacity analysis

Qwen

The Geometry of Categorical and Hierarchical Concepts in Large Language Models (2024), When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs (2026)measured

Categorical concepts are represented as simplex polytopes

Details

models: Qwen3-4B · method: Linear probing, Activation Steering (Addition), Geometric analysis, Causal inner product

Emergent Manifold Separability during Reasoning in Large Language Models (2026)measured

Concept manifolds transiently untangle right before a reasoning step

Details

Chun, Polo & Chung apply mean-field Manifold Capacity Theory to residual-stream activations of real LLMs during chain-of-thought reasoning on Boolean-logic and eligibility tasks [chun-etal-2026-emergent-manifold-separability-during-reasoning] Concept manifolds transiently untangle into highly linearly-separable configurations immediately before the tokens implementing a logical computation, then recompress afterward, a pattern the authors term Dynamic Manifold Management [chun-etal-2026-emergent-manifold-separability-during-reasoning] This transient signature diverges from a comparatively stable linear-probe-accuracy analysis over the same span, so manifold-shape measurement captures reasoning-time dynamics invisible to probe accuracy [chun-etal-2026-emergent-manifold-separability-during-reasoning] The pattern holds in Ministral 3 8B Reasoning and replicates across Qwen2.5-14B/7B-Instruct and gpt-oss-20b; purely observational [chun-etal-2026-emergent-manifold-separability-during-reasoning]

models: Qwen2.5-14B-Instruct, Qwen2.5-7B-Instruct · method: Neural manifold capacity analysis

The Geometry of Categorical and Hierarchical Concepts in Large Language Models (2024), When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs (2026)measured

Semantic hierarchy is encoded as orthogonality between concept levels

Details

models: Qwen3-4B · method: Geometric analysis, Causal inner product

Mistral

The Geometry of Categorical and Hierarchical Concepts in Large Language Models (2024), When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs (2026)measured

Categorical concepts are represented as simplex polytopes

Details

models: Mistral-7B-v0.3 · method: Linear probing, Activation Steering (Addition), Geometric analysis, Causal inner product

Emergent Manifold Separability during Reasoning in Large Language Models (2026)measured

Concept manifolds transiently untangle right before a reasoning step

Details

models: Ministral 3 8B Reasoning · method: Neural manifold capacity analysis

Uncovering Symmetry Transfer in Large Language Models via Layer-Peeled Optimization (2026)measured

Prompt-set symmetries provably transfer into simplex-ETF and circulant geometry

Details

Du, He & Su prove a symmetry-transfer principle: the symmetry group acting on next-token completions is inherited by the optimal class-mean/embedding geometry under the layer-peeled objective [du-etal-2026-uncovering-symmetry-transfer] Permutation symmetry forces a Simplex Equiangular Tight Frame while cyclic-shift symmetry forces a circulant matrix structure [du-etal-2026-uncovering-symmetry-transfer] Both predictions are confirmed on real Mistral-7B-Instruct-v0.3 output-projection and context-embedding Gram matrices for permutation- and cyclic-symmetric prompt sets (days of the week, months of the year) [du-etal-2026-uncovering-symmetry-transfer]

models: Mistral-7B-Instruct-v0.3 · method: Geometric analysis, Analytical derivation

The Geometry of Categorical and Hierarchical Concepts in Large Language Models (2024), When Language Representations Interact: Separability and Cross-Lingual Effects in LLMs (2026)measured

Semantic hierarchy is encoded as orthogonality between concept levels

Details

models: Mistral-7B-v0.3 · method: Geometric analysis, Causal inner product

GPT-OSS

Emergent Manifold Separability during Reasoning in Large Language Models (2026)measured

Concept manifolds transiently untangle right before a reasoning step

Details

models: GPT-OSS-20B · method: Neural manifold capacity analysis

SimCLR

Directional Neural Collapse Explains Few-Shot Transfer in Self-Supervised Learning (2026)measured

SSL encoders collapse variance along each task's decision axis, aiding few-shot transfer

Details

Luthra et al. extend Neural Collapse to self-supervised encoders via directional class-distance-normalized variance (directional CDNV), which measures within-class variance only along a task's own class-separating axis [luthra-etal-2026-directional-neural-collapse] Across SimCLR, VICReg, MAE, and DINOv2 encoders (trained from scratch on mini-ImageNet plus off-the-shelf checkpoints), directional CDNV falls sharply while isotropic CDNV stays large, so SSL collapses variance selectively along task-relevant axes [luthra-etal-2026-directional-neural-collapse] When this holds across many independently-sampled tasks at once, their decision axes become near-orthogonal (pairwise cosine), geometrically explaining multitask few-shot transfer [luthra-etal-2026-directional-neural-collapse] No causal intervention is performed [luthra-etal-2026-directional-neural-collapse]

models: SimCLR ResNet-18 (mini-ImageNet) · method: Directional class-distance-normalized variance (directional CDNV)

Vision Transformer (ViT)

Directional Neural Collapse Explains Few-Shot Transfer in Self-Supervised Learning (2026)measured

SSL encoders collapse variance along each task's decision axis, aiding few-shot transfer

Details

models: MAE ViT-Base (Masked Autoencoder) · method: Directional class-distance-normalized variance (directional CDNV)

DINOv2

Directional Neural Collapse Explains Few-Shot Transfer in Self-Supervised Learning (2026)measured

SSL encoders collapse variance along each task's decision axis, aiding few-shot transfer

Details

models: DINOv2 ViT-B/16 · method: Directional class-distance-normalized variance (directional CDNV)

CLIP (Contrastive Language-Image Pretraining)

Directional Neural Collapse Explains Few-Shot Transfer in Self-Supervised Learning (2026)measured

SSL encoders collapse variance along each task's decision axis, aiding few-shot transfer

Details

models: CLIP ViT-B/32 · method: Directional class-distance-normalized variance (directional CDNV)

ResNet

Hard-Negative Sampling for Contrastive Learning: Optimal Representation Geometry and Neural- vs Dimensional-Collapse (2024)measured

A real ResNet-50 trained with hard-negative contrastive losses and unit-ball normalization converges to Neural-Collapse ETF geometry on CIFAR-100, while omitting either yields Dimensional-Collapse instead

Details

A real ResNet-50 is trained from scratch for 400 epochs (Adam, batch size 512) on real CIFAR-100 (and CIFAR-10, TinyImageNet) with representation dimension set to d = C - 1 and unit-ball/unit-sphere feature normalization, under supervised and unsupervised hard-negative contrastive losses (HSCL, HUCL) [jiang-etal-2023-hard-negative-sampling-neural-dimensional-collapse] Under these conditions the trained representations converge to Neural-Collapse (NC) geometry -- class means forming an Equiangular Tight Frame, measured via the zero-sum, unit-norm, and equal-inner-product metrics matching their theoretical minima -- for a range of hardness levels and batch sizes from 64 to 512 [jiang-etal-2023-hard-negative-sampling-neural-dimensional-collapse] Without incorporating hard negatives or feature normalization, Adam optimization of the same losses on the same real ResNet-50/CIFAR-100 setup instead produces Dimensional-Collapse (DC), failing to attain the NC-ETF geometry [jiang-etal-2023-hard-negative-sampling-neural-dimensional-collapse]

models: ResNet-50 (hard-negative supervised/unsupervised contrastive, CIFAR-100) · method: Geometric analysis

Diagnosing Generalization Failures from Representational Geometry Markers (2026)measured

Penultimate-layer manifold geometry predicts OOD generalization failure

Details

Chou et al. apply the GLUE framework (Geometry Linked to Untangling Efficiency) computing effective dimension, radius, and utility of penultimate-layer object manifolds in CNN classifiers (ResNet-18/34/50, VGG-19, RegNet, MobileNet, WideResNet); no ViT is used [chou-etal-2026-diagnosing-generalization-failures] Low effective manifold dimensionality and low utility (feature overspecialization) track and precede poor out-of-distribution generalization more reliably and earlier than in-distribution accuracy [chou-etal-2026-diagnosing-generalization-failures] These in-distribution geometric markers correlate with OOD performance far more strongly than in-distribution accuracy, sparsity, or covariance measures [chou-etal-2026-diagnosing-generalization-failures] The framework builds on Chung manifold-capacity theory as a foundation but does not operationalize a single scalar capacity; the analysis is observational-diagnostic [chou-etal-2026-diagnosing-generalization-failures]

models: ResNet (image classifier, various depths) · method: GLUE (Geometry Linked to Untangling Efficiency)

How Label Imbalance Shapes Geometry: A General Spectral Analysis of Multi-Label Neural Collapse (2026)measured

A real ResNet-18 trained on multi-label CIFAR-10 obeys a class-frequency-weighted prototype synthesis rule rather than uniform averaging, with collapse metrics tracking label imbalance

Details

A real ResNet-18 is trained from scratch (SGD, batch 128, momentum 0.9, cosine-annealed learning rate 0.1 to 0.001, 200 epochs) on a real trained multi-label construction built from CIFAR-10 images (MLab-CIFAR10, 40,000 samples with per-sample label multiplicity up to 2) [ma-etal-2026-multi-label-neural-collapse-spectral-analysis] Testing whether higher-multiplicity class prototypes are the uniform average of their constituent single-label means, a Gram-residual alignment test shows the residual drops from 36.14 to 0.116 when the last-layer weight matrix is rescaled by the class-count-aware factor D^(-1/2) predicted by the paper's class-frequency-weighted synthesis rule, versus 2.33x10^5 under the wrong-direction D^(1/2) scaling [ma-etal-2026-multi-label-neural-collapse-spectral-analysis] As the class-imbalance ratio is reduced from 1.0 to 0.2 to 0.1 (epoch 200), the NC2 and NC3 Neural-Collapse collapse metrics increase monotonically, showing label imbalance systematically distorts the classical (balanced) Neural-Collapse geometry away from a symmetric Equiangular Tight Frame [ma-etal-2026-multi-label-neural-collapse-spectral-analysis]

models: ResNet-18 (multi-label, MLab-CIFAR10) · method: Geometric analysis

Prevalence of Neural Collapse during the terminal phase of deep learning training (2020), Are Neurons Actually Collapsed? On the Fine-Grained Structure in Neural Representations (2023)measured

Last-layer class means converge to a Simplex Equiangular Tight Frame

Details

Papyan, Han & Donoho show that in the terminal phase of training, standard image classifiers (VGG/ResNet/DenseNet) undergo Neural Collapse, with within-class variability collapsing (NC1) [papyan-etal-2020-neural-collapse] Class means become equinorm and maximally equiangular, converging to a Simplex Equiangular Tight Frame (NC2), classifier weights become self-dual with the class means (NC3), and the decision rule converges to nearest-class-center (NC4) [papyan-etal-2020-neural-collapse] All four are demonstrated across three architectures and seven datasets (480 models), and an information-theoretic optimal-code argument proves the Simplex ETF is the unique optimum [papyan-etal-2020-neural-collapse] Yang et al. caveat that NC1's collapse is not complete: a fine-grained information-bearing structure survives inside each cluster, and unsupervised clustering of collapsed CIFAR-10 super-class representations recovers the original 10 classes at 93% [yang-etal-2023-are-neurons-actually-collapsed]

models: ResNet (image classifier, various depths) · method: Geometric analysis, Analytical derivation

On the Geometry of Generalization and Memorization in Deep Neural Networks (2021)measured

Object-manifold radius and dimension track memorization and double descent

Details

Stephenson et al. apply mean-field manifold capacity theory to AlexNet, VGG-16, and ResNet-18 classifiers trained on CIFAR-100 and Tiny-ImageNet, tracked jointly across depth and training epochs [stephenson-etal-2021-geometry-of-generalization-and-memorization] Object-manifold radius and dimension track memorization: manifolds for memorized examples differ systematically from generalized ones [stephenson-etal-2021-geometry-of-generalization-and-memorization] Manifold dimension undergoes double descent (shown for a ResNet-18 width-sweep on CIFAR-100 with 10% label noise), while radius and center correlation stay monotonic with model size [stephenson-etal-2021-geometry-of-generalization-and-memorization] Individual-example manifold radius and dimension shrink as examples move from a memorized to a generalized regime; purely observational [stephenson-etal-2021-geometry-of-generalization-and-memorization]

models: ResNet (image classifier, various depths) · method: Neural manifold capacity analysis

VGG

Diagnosing Generalization Failures from Representational Geometry Markers (2026)measured

Penultimate-layer manifold geometry predicts OOD generalization failure

Details

models: VGG (image classifier, various depths) · method: GLUE (Geometry Linked to Untangling Efficiency)

Prevalence of Neural Collapse during the terminal phase of deep learning training (2020), Are Neurons Actually Collapsed? On the Fine-Grained Structure in Neural Representations (2023)measured

Last-layer class means converge to a Simplex Equiangular Tight Frame

Details

models: VGG (image classifier, various depths) · method: Geometric analysis, Analytical derivation

On the Geometry of Generalization and Memorization in Deep Neural Networks (2021)measured

Object-manifold radius and dimension track memorization and double descent

Details

models: VGG (image classifier, various depths) · method: Neural manifold capacity analysis

Bayesian Wind Tunnel Mamba

The Bayesian Geometry of Transformer Attention (2026)measured

A Mamba model discovers the corner geometry of the belief simplex

Details

Agarwal, Dalal & Misra train a from-scratch Mamba/SSM on HMM filtering and find its final-layer representations organize into five discrete clusters, one per hidden state [agarwal-etal-2026] The paper describes this as discovering the corner geometry of the belief simplex, concentrating at the vertices rather than the transformer's continuous interior-populated Mixed-State Presentation [agarwal-etal-2026] A permutation/relabeling test confirms the clustering tracks genuine posterior structure, not arbitrary state IDs [agarwal-etal-2026] Mamba realizes belief accumulation and belief transport but not random-access binding, distinguishing it from the transformer's richer continuous geometry [agarwal-etal-2026]

models: Bayesian Wind Tunnel Mamba (HMM filtering task, 9 layers, d_model=256, state dim 16) · method: PCA, Geometric analysis

DenseNet

Prevalence of Neural Collapse during the terminal phase of deep learning training (2020), Are Neurons Actually Collapsed? On the Fine-Grained Structure in Neural Representations (2023)measured

Last-layer class means converge to a Simplex Equiangular Tight Frame

Details

models: DenseNet (image classifier, various depths) · method: Geometric analysis, Analytical derivation

Pythia

(How) Do Language Models Track State? (2025)measured

Permutation-tracking transformers encode state via a parity axis orthogonal to a triangle

Details

Li, Guo & Andreas fine-tune Pythia-160M on the symmetric-group (S3, S5) word problem and find models converge on one of exactly two mechanisms, identified via probing and activation patching [li-etal-2025] The Associative Algorithm computes the product via a hierarchical parallel-prefix scan, with state-probe accuracy and patch-restoration length growing exponentially with depth [li-etal-2025] The Parity-Associative Algorithm first computes parity via dedicated parity heads, then refines with an AA-like scan, and an intermediate pretraining curriculum controls which mechanism emerges [li-etal-2025] For PAA on S3, PCA reveals an orthogonal parity axis crossed with a 2D triangle of three equidistant clusters (a 2-simplex direct-summed with the parity direction); the clean 3-direction decomposition breaks down for S5 [li-etal-2025]

models: Pythia-160M · method: Linear probing, Activation patching, PCA

DeepSpeech

Untangling in Invariant Speech Recognition (2019)measured

DeepSpeech2 untangles word and phoneme manifolds while tangling speaker

Details

Stephenson et al. apply mean-field manifold capacity theory to a real trained DeepSpeech2 ASR network (960 hours of LibriSpeech) for phoneme, speaker, and word manifolds [stephenson-etal-2019-untangling-in-invariant-speech-recognition] Word and phoneme manifolds progressively untangle (linear separability rises) across network depth and recurrent processing time [stephenson-etal-2019-untangling-in-invariant-speech-recognition] Speaker manifolds do the opposite: their separability decreases with depth, dropping even below the untrained network, as the network builds speaker-invariance by increasing speaker-manifold dimensionality [stephenson-etal-2019-untangling-in-invariant-speech-recognition] Word and phoneme separability also untangle temporally, peaking near the location of the relevant word [stephenson-etal-2019-untangling-in-invariant-speech-recognition]

models: DeepSpeech2 (conv + batch-norm + RNN, CTC loss, trained on LibriSpeech) · method: Neural manifold capacity analysis

AlexNet

On the Geometry of Generalization and Memorization in Deep Neural Networks (2021)measured

Object-manifold radius and dimension track memorization and double descent

Details

models: AlexNet (ImageNet image classifier, supervised) · method: Neural manifold capacity analysis

Toy Superposition Autoencoder (Elhage et al.)

Toy Models of Superposition (2022)measured

Superposed features arrange into uniform polytopes with quantized dimensionality

Details

Elhage et al. train toy ReLU-output autoencoders and find a sharp first-order phase change per feature between not-learned, dedicated-dimension, and superposition [elhage-etal-2022-toy-models-of-superposition] In superposition, features arrange into small uniform polytopes (antipodal pairs, pentagons, tetrahedra, square antiprisms) at quantized sticky fractional-dimensionality values [elhage-etal-2022-toy-models-of-superposition] Adversarial-example vulnerability rises sharply (over 3x) as superposition forms and closely tracks the reciprocal of feature dimensionality [elhage-etal-2022-toy-models-of-superposition] Adversarial training causally reduces superposition, eliminating it entirely only at unreasonably large perturbation budgets (80% input L2 norm) [elhage-etal-2022-toy-models-of-superposition]

models: Toy ReLU-output autoencoder (h=Wx, x'=ReLU(W^T h + b)) · method: Analytical derivation, Geometric analysis

VAST

A TRIANGLE Enables Multimodal Alignment Beyond Cosine Similarity (2025)measured

A triangle-area tri-modal similarity loss improves retrieval

Details

Cicchetti et al. view three unit-norm modality embeddings as a 2-simplex whose closed-form area is an exact anchor-free, fusion-free measure of three-way alignment [cicchetti-etal-2025-triangle-multimodal-alignment] Replacing pairwise cosine similarity in a contrastive loss with the negative triangle area (the TRIANGLE loss) improves zero-shot video-text Recall@1 by +3.6 to +8.8 over VAST at matched capacity, and audio retrieval by up to +5.2 Recall@1 [cicchetti-etal-2025-triangle-multimodal-alignment] In a controlled toy three-modality setting TRIANGLE converges up to 4x faster than a cosine anchor loss and than the volume-based GRAM loss, but on AudioCaps GRAM outperforms TRIANGLE, so the "beats volume-based" claim is scoped to the toy setting [cicchetti-etal-2025-triangle-multimodal-alignment] TRIANGLE is a training-loss objective rather than an emergent discovered structure, with negligible compute overhead [cicchetti-etal-2025-triangle-multimodal-alignment]

models: VAST (Vision-Audio-Subtitle-Text foundation model), TRIANGLE (VAST backbone, fusion layers removed, retrained with triangle-area loss) · method: Triangle-area tri-modal similarity

Custom Behavior-Cloning Vision Encoder

Control-oriented Clustering of Visual Latent Representation (2024)measured

Control-policy visual features cluster by relative-pose orthant (Neural Collapse)

Details

Qi, Yin & Yang study behavior-cloned image-based control (no explicit label set) and define implicit classes from the sign pattern (orthant) of a relative-pose vector [qi-etal-2024-control-oriented-clustering] Across four architectures, visual representations prevalently cluster by these relative-pose-orthant classes, a control-oriented analogue of Neural Collapse discovered rather than designed [qi-etal-2024-control-oriented-clustering] Pretraining the vision encoder with a Neural-Collapse-style regularizer that encourages orthant clustering, then finetuning end-to-end, causally improves low-data control performance by 10-35%, confirmed in real-world planar pushing [qi-etal-2024-control-oriented-clustering]

models: Custom CNN vision encoder for image-based control (behavior cloning, Lunar Lander / planar pushing / block stacking) · method: Relative-pose-orthant clustering (control-oriented Neural Collapse)

Custom Research Transformer (small, purpose-built for interpretability studies)

Linguistic Collapse: Neural Collapse in (Large) Language Models (2024)measured

Neural-Collapse signatures emerge in causal language models, but not self-duality

Details

Wu & Papyan train about 30 GPT-Neo-based causal transformers from scratch on TinyStories (GPT-2 used only for its tokenizer), reframing next-token prediction as extreme imbalanced classification and measuring Neural Collapse [wu-papyan-2024-linguistic-collapse] Within-class variability collapse (NC1), the GNC2 hyperspherical-uniformity variant, uniform duality (UNC3), and NC4 strengthen with model width, depth, and training [wu-papyan-2024-linguistic-collapse] Self-duality (NC3) does NOT develop with scale, and a true Simplex ETF is unreachable because the number of classes far exceeds the hidden dimension (C >> d+1), so the geometry only tends toward it [wu-papyan-2024-linguistic-collapse] This is the first systematic demonstration of Neural-Collapse-style geometry in real, generatively pretrained causal language models [wu-papyan-2024-linguistic-collapse]

models: Custom GPT-2-style causal transformer (205M, TinyStories, largest of a 3.4M-205M width/depth/epoch grid) · method: Geometric analysis