MATH · IN · MODELS

Anisotropy

PROPERTYdistribution-propertyintermediatehow it's classified →

A representation space is anisotropic if activation vectors are not directionally uniform — they occupy a narrow angular cone rather than pointing every which way, so that two randomly chosen vectors have unexpectedly high cosine similarity on average. Origin: Ethayarajh (2019), building on Mimno & Thompson (2017)'s term for static skip-gram embeddings.

Replicationcomputed from the corpus — never hand-assigned
33 papers · no shared authors11 architecture classes · across papers3 domains · across papers44 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

A set of vectors X={x1,,xn}RdX=\{x_1,\dots,x_n\}\subset\mathbb{R}^d is isotropic if it is directionally uniform — informally, no direction is preferred over any other. Ethayarajh (2019) operationalizes the degree of anisotropy of a representation space (e.g. one layer of a contextualizing model) via the expected cosine similarity between two uniformly-randomly-sampled representations from that space:

Aniso(f)=Ex,yU(O)[cos(f(x),f(y))]\mathrm{Aniso}(f_\ell) = \mathbb{E}_{x,y\sim U(O)}\big[\cos(f_\ell(x), f_\ell(y))\big]

where OO is the set of all word occurrences and ff_\ell maps an occurrence to its representation in layer \ell. This statistic lies in [1,1][-1,1]. A value near 11 indicates one specific, one-sided form of extreme anisotropy — almost every pair of vectors points in nearly the same direction (a single narrow cone).

The mean cosine is one particular, one-sided probe — not a full isotropy test. Two cautions:

  • Aniso0\mathrm{Aniso}\approx 0 does not imply isotropy. Counterexample: put half the unit mass along +v+v and half along v-v. The cloud is maximally anisotropic (all mass on one line), yet E[cos(x,y)]=0\mathbb{E}[\cos(x,y)] = 0 because same-sign and opposite-sign pairs cancel. A near-zero mean cosine only rules out a single dominant cone, not bimodal/multimodal concentration.
  • “Isotropy” is several inequivalent notions. Distinguish at least: spherical symmetry of the distribution; isotropic covariance Cov(X)=σ2I\mathrm{Cov}(X)=\sigma^2 I; uniformity of the normalized directions on the sphere; a flat covariance spectrum; and absence of a dominant mean direction. These do not coincide without extra assumptions.

Centering is essential. A high mean cosine often reflects a large mean μ=E[X]\mu=\mathbb{E}[X] rather than covariance anisotropy. It is better to report several quantities separately: the mean-direction norm Amean=1nixi/xiA_{\text{mean}}=\big\|\tfrac1n\sum_i x_i/\|x_i\|\big\|, the spectral anisotropy Aspec=λmax(Σ^)/(trΣ^/d)A_{\text{spec}}=\lambda_{\max}(\widehat\Sigma)\big/(\operatorname{tr}\widehat\Sigma/d) of the centered data, and the angular distribution of normalized vectors — keeping the raw mean pairwise cosine as one specific empirical indicator, not the definition of (an)isotropy.

Intuition

Picture every activation vector as an arrow from the origin. In an isotropic space, the arrows point every which way, roughly evenly covering all directions — like the spines of a sea urchin. In an anisotropic space, almost all the arrows are bunched into a narrow ice-cream-cone-shaped wedge around some common direction, even though they can still have very different lengths and slightly different angles within that wedge.

Properties

  • A property of the whole representation space, not of any one vector’s meaning. Anisotropy is a statistic about how a population of vectors is directionally distributed — it says nothing on its own about whether any individual vector’s direction is semantically meaningful (contrast with Linear Direction, which is about one specific direction’s causal/semantic role).

  • Must be adjusted for before other geometric statistics are trusted. Ethayarajh (2019) shows the same-word self-similarity and same-sentence intra-sentence similarity statistics are both inflated by ambient anisotropy — a word’s representations can look “well contextualized” or “poorly contextualized” purely as an artifact of how anisotropic the whole layer is, unless the raw statistic has the layer’s own anisotropy baseline subtracted off first.

  • Empirically present and increasing with depth in every contextualizing model tested (ELMo, BERT, GPT-2). Ethayarajh (2019) finds all three models’ contextualized layers are anisotropic (the only isotropic layer is ELMo’s non-contextual, character-level input layer), generally worsening in upper layers — in GPT-2’s last layer, two randomly chosen words have on average almost perfect cosine similarity.

  • In tension with known benefits of isotropy for static embeddings. Arora et al. (2017) show isotropy aids theoretical self-normalization; Mu et al. (2018) show empirically that post-hoc isotropizing static embeddings (mean-subtraction) improves downstream performance — making the degree of anisotropy found in contextualized representations “surprising” against that prior expectation, not an obviously-benign byproduct.

  • A cone in the directional sense here is a statistical/global property of the whole point cloud, not the same object as Cone. manifolds-cone’s polyhedral/circular cone catalogues a specific, causally-meaningful set of generator directions (e.g. a behavior-mediating cone spanned by a handful of steering vectors); anisotropy’s “narrow cone” is a global statistical description of an entire layer’s activation cloud, with no claim that the shared direction it clusters around is itself meaningful.

  • Not actually diffuse across many dimensions — dominated by a handful of “rogue” outliers. Timkey & van Schijndel (2021) decompose the cosine-similarity anisotropy statistic per-dimension and find just 1-5 high-variance, far-from-origin dimensions account for nearly all of it (in the most extreme case, one XLNet dimension explains over 99%); via KL-divergence ablation they further show these cosine-dominant dimensions are largely not the dimensions the model’s own behavior actually depends on — a “striking mismatch” between what dominates the geometric statistic and what matters causally. This reframes anisotropy as substantially an artifact of a small number of outlier dimensions rather than a genuinely global directional bias, and motivates per-dimension standardization (z-scoring) as a correction before trusting any cosine-similarity-based geometric analysis — see rogue-dimensions-obscure-anisotropy.

  • IsoScore-family metrics extend beyond static contextualization studies to a training-dynamics setting, where forcibly restoring isotropy turns out to hurt rather than help. Lanza, Pereira, Miozzo, Angelats & Dini (2026) measure IsoScore and a generalized IsoEntropy (covariance-eigenvalue-spectrum uniformity) on penultimate-layer features of real ResNet-18 encoders trained via SupCon, Co2L, and two neural-collapse-inspired variants on CIFAR-10/100 under class-incremental continual learning, finding isotropy declines as more tasks accumulate but is not consistently correlated with accuracy across methods. A causal test — adding a differentiable isotropy regularizer to the training loss — raises IsoScore sharply (0.023 to 0.969 on one config) while collapsing accuracy (70.64% to 14.74%), showing that forcibly restoring isotropy in a continual-learning setting is actively harmful rather than a free fix, a cautionary causal counterpoint to isotropy-improving interventions found beneficial elsewhere. See lanza-etal-2026-real-resnet18-feature-spaces-lose-isotropy-under-continual-learning-and-forcing-isotropy-back-up-via-a-regularizer-degrades-rather-than-helps-accuracy.

  • The measurement instrument itself matters, and a rigorously axiomatized one gives a different picture again. Rudman et al. (2021)‘s IsoScore — built to be mean-agnostic, scale-invariant, rotation-invariant, and monotonically sensitive to how many dimensions are used, properties average-cosine-similarity and the “partition score” each provably lack — finds contextualized models are even more anisotropic than cosine-based estimates suggested (GPT/GPT-2 don’t uniformly occupy even one full dimension), yet directly overturns two conclusions the older metrics had supported: there is no significant trend of increasing isotropy in earlier layers, and mean-centering does not “recover” isotropy (isotropy here is a function of the covariance matrix alone, unaffected by the mean by construction) — see isoscore-refines-anisotropy-picture.

  • Substantially, though not entirely, a between-cluster effect. Cai, Huang, Bian & Church (2021) cluster each layer’s representations (k-means) and recompute cosine-similarity statistics within each cluster after center-shifting: inter-type cosine similarity is close to 00 (isotropic) within clusters across every layer of BERT, DistilBERT, GPT, GPT-2, and ELMo, even though the unadjusted, whole-population statistic reads as strongly anisotropic — reframing (not overturning) Ethayarajh’s finding as real at the population level but substantially explained by distinct clusters occupying different regions of the space, rather than every individual vector pointing in a similar direction. The same paper separately finds a low-dimensional manifold structure (via Local Intrinsic Dimension) coexisting with this within-cluster directional isotropy — two distinct senses of “isotropy” that are both true at once, not in tension. See clusters-explain-apparent-anisotropy and Intrinsic-dimension profile across depth‘s manifold-dimension-lower-than-ambient.

  • Frequency-conditioned, and correctable without discarding information. Li et al. (2020) show BERT’s anisotropic word/sentence embedding space is not uniformly distorted but specifically frequency-biased — high-frequency words cluster near the origin, low-frequency words sit far away and disperse sparsely among their own neighbors, leaving low-probability-density “holes” a sentence embedding (an average of context embeddings) can land in. BERT-flow maps the embedding space to an isotropic Gaussian via an invertible, information-preserving normalizing flow, correcting this without the information loss of nulling top singular vectors — see bert-flow-corrects-anisotropic-holes.

  • A much simpler closed-form linear fix works nearly as well as a learned one. Huang et al. (2021) show whitening — an exact, trainable-parameter-free linear map computed directly from the embedding covariance’s eigendecomposition — consistently improves sentence-similarity performance across a wide range of pretrained models, without the training cost of a full normalizing flow, since matching just the first two empirical moments (mean and covariance) to an isotropic target already captures most of the practical benefit — see whitening-improves-sentence-similarity.

  • The modality gap is a mode-failure of the InfoNCE loss specifically at low temperature, not primarily an initialization/cone effect, and adding intra-modal negative pairs (xNCE) reduces the gap while matching or improving retrieval and zero-shot accuracy. Mager, Nassar & Hansen (2026) prove and empirically demonstrate this in a uni-modal control setting (two independently-initialized encoders under identical initialization conditions) with CLIP-style dual encoders, isolating the loss function’s own contribution to the gap from initialization dynamics — a third competing/complementary account of the modality gap’s cause alongside the cone-effect account and the loss-uniformity account already in this section. See the-clip-modality-gap-is-an-infonce-low-temperature-mode-failure-fixed-by-adding-intra-modal-negatives.

  • Under a subspace constraint (dimension collapse), the modality gap provably converges to the angle between two modality-specific hyperplanes, and closing that angle via a shared-space projection on real CLIP causally improves both the gap and zero-shot accuracy over prior baselines. Yi, Douady & Chen (2025) prove the gap converges to zero without a subspace constraint or under a cone constraint, but to the smallest angle between two hyperplanes once dimension collapse is imposed, then apply their own Shared Space Projection to real CLIP ViT-B/32 embeddings, closing the gap angle from 74.69 to 5.37 degrees on CIFAR-10 (vs. 7.02 for a translation baseline) while better preserving zero-shot retrieval accuracy (86.43 vs. 80.97) — a fourth account, the only one with a formal necessary-and-sufficient proof tying dimension collapse to a nonzero gap. See the-clip-modality-gap-converges-to-the-angle-between-two-modality-specific-hyperplanes-under-dimension-collapse.

  • A per-token geometric signature combining cross-layer covariance dimensionality and circular-variance anisotropy detects LLM hallucination better than prior uncertainty baselines. Srey et al. (2026) show their SIVR signature (log-pseudo-determinant of cross-layer hidden-state covariance plus circular variance of layer-wise normalized hidden vectors) reaches up to 89.5% AUC on TriviaQA for Llama-3.1-8B, beating prior baselines across seven QA/reasoning benchmarks. See per-token-log-pseudo-determinant-of-cross-layer-covariance-and-circular-variance-of-hidden-states-detect-llm-hallucination-better-than-prior-uncertainty-baselines.

  • The residual stream’s distributional skewness is measurably higher in later layers when an LLM favors contextual over parametric knowledge under a conflict. Zhao et al. (2024) quantify this via Kurtosis/Hoyer/Gini metrics from layers 20-30 in Llama3-8B and Llama2-7B, a directly measured geometric-shape difference tied to which knowledge source the model relies on. See logistic-regression-probes-on-the-residual-stream-detect-parametric-vs-contextual-knowledge-conflict-and-predict-which-source-the-model-will-rely-on-before-generation-with-a-distinct-skewness-signature.

  • The fraction of embedding variance held by the single dominant (“rogue”) dimension predicts whether cosine or rank-based similarity metrics win, across 19 text encoders, and causally projecting out dominant dimensions erases the effect only where it was anisotropic to begin with. Parupudi (2026) finds rank correlation 0.86 and linear correlation 0.95 between rogue-dimension variance share and the metric-improvement magnitude. See the-fraction-of-embedding-variance-held-by-the-single-dominant-dimension-predicts-whether-cosine-or-rank-based-similarity-metrics-win-across-19-text-encoders-and-projecting-out-dominant-dims-erases-the-effect.

  • Foundational: static word embeddings share a dominant mean plus a handful of top-PC directions, and removing them causally improves downstream tasks. Mu & Viswanath (2018) measure, via PCA/SVD on word2vec/GloVe-family embedding matrices across multiple languages, that a large shared mean vector plus a small number of top principal directions account for most of the variance — a foundational, pre-contextualized-model anisotropy measurement. Their “all-but-the-top” postprocessing (subtract the mean, project out the top DD directions) causally improves word similarity, analogy, categorization, STS, and text-classification benchmarks with no retraining, establishing both the measurement and the causal-correction pattern that later contextualized-model anisotropy work (Ethayarajh 2019 onward) builds on. See mu-viswanath-2018-word-embeddings-share-a-dominant-mean-vector-plus-top-pc-anisotropy-and-removing-them-causally-improves-downstream-task-performance.

  • Foundational: tied input-output embeddings in real trained NLG models degenerate into a narrow cone, with a training-dynamics explanation for why. Gao, He, Tan, Qin, Wang & Liu (2019) measure this cone directly in real trained language/translation models with weight-tied embeddings and explain it as a consequence of softmax normalization combined with weight tying, which pushes especially rare-word vectors toward a shared direction; their proposed cosine-similarity regularizer (MLE-CosReg, not “MMI”) measurably widens the embedding spread. Predates and matches the pattern later confirmed in contextualized (BERT-family) models. See gao-etal-2019-tied-input-output-embeddings-in-real-trained-nlg-models-degenerate-into-a-narrow-anisotropic-cone-explained-by-softmax-weight-tying-dynamics.

  • A mixture-of-experts router suppresses its own hidden states’ dominant shared direction under load-balancing loss, and hidden-state similarity predicts expert-usage similarity across five real production MoE models. Wang, Hayou & Nalisnick (2026) show real hidden states in five production MoE models — gpt-oss-20B, ERNIE-4.5-21B-A3B-Base, Qwen-3-30B, Ling-mini and Trinity-Mini-Base (Moonlight-16B is a comparison only) — develop a dominant shared direction (first right singular vector) whose retained energy grows with depth, and prove the auxiliary load-balancing loss forces the router to suppress this shared direction to maintain diverse outputs; separately, hidden-state cosine similarity tracks expert-usage similarity so closely that two different models’ solutions to the same question show ~60% expert overlap, comparable to the same model on two different questions. See wang-etal-2026-real-pretrained-moe-routers-suppress-a-shared-dominant-hidden-state-direction-under-load-balancing-loss-and-hidden-state-cosine-similarity-predicts-expert-usage-similarity-across-five-production-models.

  • Visual tokens fed into real trained MLLMs partition into discrete functional clusters, each collapsing to near-perfect cross-image cosine similarity. Fan, Tong, Zhao & Shen (2026) cluster LLaVA-1.5-7B’s visual tokens (norm threshold plus cosine similarity of cluster centroids) into “Sink,” “Dead,” and “Alive” categories (sink+dead ~40% of all tokens); sink-cluster centroids reach cross-image cosine similarity >0.99 (ViT stage) and >0.95 (LLM stage) — a near-total directional collapse for specific functional subgroups rather than the whole population, confirmed causally distinct via cluster-targeted ablation (removing sink/dead tokens leaves overall accuracy flat or improves a CV-centric sub-score). See fan-etal-2026-visual-tokens-fed-into-real-trained-mllms-partition-into-sink-dead-and-alive-functional-clusters-with-near-perfect-cross-image-cosine-similarity-of-cluster-centroids.

  • A bell-shaped, middle-layer-peaked anisotropy profile distinguishes decoder-only from encoder architectures. Razzhigaev, Mikhalchuk, Goncharova, Oseledets, Dimitrov & Kuznetsov (2024) compute an SVD-based anisotropy score layer-by-layer across a wide range of real pretrained encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2, GPT-2, GPT-J, Falcon, Bloom, Pythia, TinyLlama): “decoders showcase a unique bell-shaped structure, indicating that the middle layers tend to have a higher anisotropy concentration,” a depth-wise shape distinct from encoders’ own profile — purely observational, no causal intervention. See razzhigaev-etal-2024-decoder-only-transformers-develop-a-bell-shaped-middle-layer-anisotropy-concentration-unlike-encoders-across-a-wide-range-of-real-pretrained-models.

  • “Massive values” concentrate at consistent, cross-head dimension indices in real RoPE transformers’ query/key space, absent without RoPE. Jin, Mei, Xu, Sun, Tang, Du, Liu & Zhang (2025, ICML 2025) define a massive value as a per-(head, dimension) L2 norm exceeding 5x the mean, finding these cluster at nearly identical dimension indices across attention heads in real Llama-3-8B-Instruct, Gemma-2-9B-it, Qwen2.5-7B-Instruct and Llama-2-7B-chat, but not in non-RoPE GPT-2 or OPT. Ablating (mean/zero/min replacement) the massive-value dimensions during prefilling collapses contextual-knowledge tasks (e.g. GSM8K 76.9%→4.0% on Llama3-8B; passkey retrieval 100%→0%), and quantization schemes that preserve these dimensions (AWQ/SmoothQuant) retain performance that GPTQ (which does not) loses — a rogue/outlier-dimension-style concentration pattern with a strong causal signature, distinct in mechanism (tied specifically to rotary position embeddings) from the cosine-similarity-decomposition route of Timkey & van Schijndel above. See jin-etal-2025-massive-values-concentrate-at-consistent-head-dim-indices-in-real-rope-transformers-q-k-space-and-ablating-or-quantizing-them-collapses-contextual-knowledge-tasks.

  • Positional encodings causally anchor an index-based, rather than content-based, spatial-similarity organization in a real trained ViT. Mannes (2026) computes Spatial Similarity Distance Correlation (SSDC — the Spearman correlation between representation cosine similarity and negative spatial distance) across depth in a real ViT-S/16 trained from scratch on ImageNet-100, across five positional-encoding variants; all PE schemes shift token representations toward index-anchored rather than purely content-based spatial organization, and causal interventions (random permutation, positional-embedding magnitude scaling) show SSDC collapsing alongside rising fragility to distribution shift as positional-embedding magnitude is scaled down — a distance-dependent spatial-organization claim in a new (vision) modality, structurally analogous to Zuo, Guerzhoy & Guerzhoy’s distance-decaying adjacency pattern above. See mannes-2026-positional-encodings-causally-anchor-an-index-based-rather-than-content-based-spatial-similarity-organization-in-a-real-trained-vit-and-removing-it-increases-fragility.

Exercises

Base

  1. If every vector in XX were identical (all equal to some fixed v0v\neq 0), what would Aniso(f)\mathrm{Aniso}(f_\ell) equal?
Solution

cos(v,v)=1\cos(v,v)=1 for every pair, so Aniso(f)=1\mathrm{Aniso}(f_\ell)=1 — the maximum possible anisotropy, consistent with a “cone” that has collapsed to a single ray.

  1. Why does Ethayarajh (2019) subtract a per-layer anisotropy baseline from self-similarity and intra-sentence similarity before interpreting them, rather than using the raw statistics directly?
Solution

Because a high raw self-similarity could mean either “this word’s representations are genuinely stable across contexts” or simply “this whole layer is so anisotropic that even two unrelated random words would already have high cosine similarity” — the two explanations are observationally identical unless the layer-wide baseline (average similarity between random word pairs) is subtracted off first, isolating the word-specific signal from the ambient directional bunching.

Middle

  1. Ethayarajh reports that GPT-2’s raw (non-anisotropy-adjusted) maximum explainable variance (MEV) is around 30% in middle layers, while the anisotropy-adjusted MEV stays under 5% in every layer of every model tested. Explain how both numbers can be true simultaneously without contradiction.
Solution

Raw MEV measures how much variance a word’s own representations’ first principal component explains, without accounting for the fact that a single shared direction (the anisotropic bulk direction) already explains a large fraction of variance across any random collection of vectors in that layer, word-specific or not. Subtracting the baseline MEV (computed the same way but over randomly sampled, unrelated word representations) isolates how much additional, word-specific variance-explaining power the first principal component has beyond what ambient anisotropy alone would already provide. GPT-2’s high raw MEV is therefore mostly attributable to its extreme anisotropy, not to genuine word-sense clustering — exactly the adjustment’s purpose.

  • The same “rogue dimension” phenomenon, extended to a new modality and given a quantified positional structure. Cheng, Zheng, Xie, Liao & Li (2026) show video diffusion transformers (Wan2.1, Wan2.2, CogVideoX) have their own massive-activation channels — a small, fixed set of dimensions taking magnitudes “often exceeding 50x the mean activation value,” consistent across layers, model scales, and text conditioning. Unlike the token-agnostic outlier dimensions of Timkey & van Schijndel’s LLM setting, these follow a quantified token-position hierarchy tied to the model’s own latent temporal chunking (first-frame > latent-frame-boundary > interior tokens, with the boundary-to-interior ratio declining toward 1 over the course of denoising). Rather than only measuring the outlier dimensions’ causal (ir)relevance via ablation, Structured Activation Steering (STAS) directly overwrites their values toward a scaled reference magnitude at the identified token positions, causally improving generation quality (VBench total score, all three models) — one of the only entries in this map where amplifying, rather than removing or standardizing, an anisotropy-driving outlier dimension is shown to be the behaviorally beneficial intervention, and a uniform (position-agnostic) version of the same edit is shown to hurt quality by way of contrast. See cheng-etal-2026-massive-activation-channels-in-video-dits-follow-a-quantified-token-position-hierarchy-and-causally-steering-them-improves-video-quality.

  • A distance-dependent sub-structure distinct from the population-level statistic. Zuo, Guerzhoy & Guerzhoy (2025) show that in causal-attention Transformers with no positional encoding, the self-cosine-similarity matrix isn’t just uniformly high (the global anisotropy story) — it decays specifically with token distance: nearby tokens are systematically more similar than far-apart ones (the “adjacency pattern”), quantified by a novel row-wise monotonicity statistic (the adjacency probability score) that jumps from ~0.5 at the raw token-embedding layer to ≥0.8 immediately after the first causal-attention layer, in both randomly initialized and trained 6-layer NanoGPT models on four synthetic position-sensitive tasks. This pattern is provably expected from the approximate orthogonality of high-dimensional embeddings plus the convex-combination structure of causal-attention outputs, and using cosine similarity to the last-position embedding as a linear-probe feature predicts token position better than the embedding’s own norm/variance (a competing account of how positional information is stored). See zuo-etal-2025-causal-attention-without-positional-encodings-produces-a-distance-decaying-adjacency-pattern-in-the-self-cosine-similarity-matrix-quantified-by-a-novel-adjacency-probability-score.

  • A state-space (Mamba) architecture shows anisotropy roughly an order of magnitude tighter than BERT’s, co-occurring with representational collapse, and a geometry-targeted causal fix fails. Wagh & Singh (2026) measure a mean pairwise cosine similarity of 0.9999 in frozen Mamba-130m (vs. BERT’s ~0.99), alongside a CoLA linear probe collapsing to majority-class prediction (MCC=0.000); a modified SSM recurrence forcing new state writes to be perpendicular to the current state (“orthogonal injection”) leaves the cosine similarity unchanged and worsens STS-B correlation — a clean negative causal result reported candidly. See wagh-singh-2026-frozen-mamba-130m-activations-collapse-into-an-extremely-narrow-anisotropy-cone-10x-tighter-than-bert-and-a-geometry-targeted-orthogonal-injection-fails-to-widen-it.

  • A cross-modal magnitude asymmetry, not the usual within-modality directional cone. Qi, Liu, Tang & Zhu (2025) measure real LLaVA-1.5-7B vision- and text-token embeddings on 5,000 COCO pairs, finding vision-token L2 norms (~10^1-10^3) dwarf text-token norms (~10^-1-10^0) by a measured factor of 10-100x, suppressing the LLM’s positional (RoPE) sensitivity and causing spatial-reasoning failures — a magnitude-scale asymmetry between modalities rather than a within-modality directional cone, but structurally analogous to the CLIP/contrastive modality gap already on this page. RMS-renormalizing vision embeddings to the text-embedding norm scale causally and predictably restores positional sensitivity, improving accuracy on five benchmarks (e.g. a synthetic 2D spatial task: 56.63%→64.80% combined with multilayer CLIP-feature extraction). See llava-vision-token-embeddings-have-100x-larger-norms-than-text-suppressing-positional-sensitivity-and-normalization-restores-it.

Key papers

  • Mu, Jiaqi & Viswanath, Pramod (2018). All-but-the-Top: Simple and Effective Postprocessing for Word Representations. ICLR. arXiv:1702.01417 — foundational static-embedding anisotropy measurement and causal all-but-the-top correction.
  • Wagh, Bhagyashree & Singh, Akash (2026). Lost in State Space: Probing Frozen Mamba Representations. arXiv:2605.00253 — extreme SSM anisotropy and a negative orthogonal-injection causal test.
  • Zuo, Guerzhoy & Guerzhoy (2025). Position Information Emerges in Causal Transformers Without Positional Encodings via Similarity of Nearby Embeddings. COLING 2025, arXiv:2501.00073 — the distance-decaying “adjacency pattern” and adjacency probability score.
  • Ethayarajh (2019). How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings. EMNLP-IJCNLP, aclanthology D19-1006 — origin of the anisotropy-adjusted contextuality measures (self-similarity, intra-sentence similarity, maximum explainable variance) for contextualized LMs.
  • Mimno & Thompson (2017). The Strange Geometry of Skip-gram with Negative Sampling. EMNLP — origin of the “narrow cone” anisotropy finding for static word embeddings, which Ethayarajh (2019) extends to contextualized models.
  • Arora, Liang & Ma (2017). A Simple but Tough-to-Beat Baseline for Sentence Embeddings. ICLR — theoretical grounding for why isotropy corresponds to near-zero expected cosine similarity between random vectors.
  • Timkey & van Schijndel (2021). All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality. EMNLP, 2021.emnlp-main.372, arXiv:2109.04404 — shows anisotropy is dominated by a handful of outlier dimensions largely irrelevant to model behavior; see Rogue-dimension analysis.
  • Rudman, Gillman, Rayne & Eickhoff (2021). IsoScore: Measuring the Uniformity of Embedding Space Utilization. arXiv:2108.07344 — a rigorously axiomatized isotropy metric that overturns two prior cosine-based conclusions about layer depth and mean-centering; see IsoScore.
  • Li, Zhou, He, Wang, Yang & Li (2020). On the Sentence Embeddings from Pre-trained Language Models. EMNLP, 2020.emnlp-main.733, arXiv:2011.05864 — traces anisotropy to frequency-conditioned “holes” in the embedding space and corrects them via an invertible normalizing flow; see Normalizing-flow calibration (BERT-flow).
  • Huang, Tang, Zhong, Lu, Shou, Gong, Jiang & Duan (2021). WhiteningBERT: An Easy Unsupervised Sentence Embedding Approach. Findings of EMNLP, 2021.findings-emnlp.23 — shows a simple closed-form linear whitening transform corrects anisotropy nearly as effectively as a learned normalizing flow, across a wide range of pretrained models; see Whitening transformation.
  • Liang, Zhang, Kwon, Yeung & Zou (2022). Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning. NeurIPS 2022, arXiv:2203.02053 — extends the narrow-cone finding to two-encoder contrastive (image-text) models, proves an initialization-driven cause via two theorems, and causally manipulates the resulting inter-modality displacement vector; see clip-modality-gap-is-a-provable-cone-effect-plus-causally-tunable-offset.

Found in (33 observations · 44 families)

ELMo

How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings (2019)measured

Contextualized embeddings occupy a narrow anisotropic cone

Details

- Across BERT, ELMo and GPT-2, contextualized representations occupy a narrow anisotropic cone: two random words have high average cosine similarity, near 1.0 in GPT-2's last layer. [ethayarajh-2019] - Anisotropy increases in upper layers together with context-specificity, and self-similarity falls monotonically with depth in all three models. [ethayarajh-2019] - Fewer than 5% of a word's contextual variance is captured by a single static embedding (anisotropy-adjusted MEV). [ethayarajh-2019] - Tested on BERT-base-cased, the original 2-layer biLSTM ELMo, and GPT-2 small; analysis is observational. [ethayarajh-2019]

models: ELMo (AllenNLP biLM, 1B Word Benchmark) · method: Contextuality metrics (self-similarity, intra-sentence similarity, MEV), PCA
Isotropy in the Contextual Embedding Space: Clusters and Manifolds (2021)measured

Clusters and low-dim manifolds explain apparent anisotropy

Details

- Apparent global anisotropy (inter-type cosine ~0.7) is largely an artifact of a few isolated clusters: after k-means clustering and per-cluster mean-shifting, within-cluster cosine isotropy is near zero across every layer. [cai-etal-2021] - Contextual embeddings sit on low intrinsic-dimension manifolds, with Local Intrinsic Dimension 5.6 (BERT), 7.3 (DistilBERT), 6.8 (GPT), 7.0 (GPT-2), 9.1 (ELMo), far below the 768/1024 ambient dimensions. [cai-etal-2021] - GPT and GPT-2 additionally show a Swiss-Roll manifold tied to word frequency. [cai-etal-2021] - Tested on uncased base BERT, DistilBERT, GPT and GPT-2 (Hugging Face) and AllenNLP ELMo (no 5.5B variant); observational. [cai-etal-2021]

models: ELMo (AllenNLP biLM, 1B Word Benchmark) · method: Cluster-corrected isotropy analysis

BERT

How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings (2019)measured

Contextualized embeddings occupy a narrow anisotropic cone

Details

- Across BERT, ELMo and GPT-2, contextualized representations occupy a narrow anisotropic cone: two random words have high average cosine similarity, near 1.0 in GPT-2's last layer. [ethayarajh-2019] - Anisotropy increases in upper layers together with context-specificity, and self-similarity falls monotonically with depth in all three models. [ethayarajh-2019] - Fewer than 5% of a word's contextual variance is captured by a single static embedding (anisotropy-adjusted MEV). [ethayarajh-2019] - Tested on BERT-base-cased, the original 2-layer biLSTM ELMo, and GPT-2 small; analysis is observational. [ethayarajh-2019]

models: BERT-base-cased · method: Contextuality metrics (self-similarity, intra-sentence similarity, MEV), PCA
On the Sentence Embeddings from Pre-trained Language Models (2020)measured

Normalizing flow removes BERT's anisotropic frequency holes

Details

- BERT's sentence-embedding space is anisotropic and frequency-biased: high-frequency words sit near the origin (norm ~0.95) and low-frequency words far out (~1.45), leaving sparse low-density "holes". [li-etal-2020-bertflow] - A Glow-style invertible normalizing flow maps the frozen embeddings to an isotropic standard Gaussian, correcting anisotropy without discarding directions. [li-etal-2020-bertflow] - Flow calibration raises average STS Spearman from 62.89 to 70.98 (BERT-base) and 66.57 to 74.73 (BERT-large) without NLI supervision. [li-etal-2020-bertflow] - Tested on BERT-base and BERT-large (paper writes "BERT_base"/"BERT_large"; uncased per released code); SBERT numbers are cited baselines only. [li-etal-2020-bertflow]

models: BERT-base-uncased, BERT-large-uncased · method: Normalizing-flow calibration (BERT-flow)
Isotropy in the Contextual Embedding Space: Clusters and Manifolds (2021)measured

Clusters and low-dim manifolds explain apparent anisotropy

Details

- Apparent global anisotropy (inter-type cosine ~0.7) is largely an artifact of a few isolated clusters: after k-means clustering and per-cluster mean-shifting, within-cluster cosine isotropy is near zero across every layer. [cai-etal-2021] - Contextual embeddings sit on low intrinsic-dimension manifolds, with Local Intrinsic Dimension 5.6 (BERT), 7.3 (DistilBERT), 6.8 (GPT), 7.0 (GPT-2), 9.1 (ELMo), far below the 768/1024 ambient dimensions. [cai-etal-2021] - GPT and GPT-2 additionally show a Swiss-Roll manifold tied to word frequency. [cai-etal-2021] - Tested on uncased base BERT, DistilBERT, GPT and GPT-2 (Hugging Face) and AllenNLP ELMo (no 5.5B variant); observational. [cai-etal-2021]

models: BERT-base-uncased · method: Cluster-corrected isotropy analysis
IsoScore: Measuring the Uniformity of Embedding Space Utilization (2021)measured

IsoScore shows contextual spaces use even fewer dimensions

Details

- IsoScore measures isotropy from the covariance alone (PCA-reorient, length-normalize the variance vector, map to [0,1]), making it mean-agnostic, scale- and rotation-invariant. [rudman-etal-2021] - Every contextualized model scores below 0.18 (GPT and GPT-2 do not isotropically use even one dimension), so they use fewer dimensions than cosine-based estimates implied. [rudman-etal-2021] - IsoScore overturns two prior conclusions: isotropy does not increase in earlier layers, and mean-centering does not recover isotropy (a rotation test holds IsoScore at 0.216 while average cosine swings 0.97-0.99). [rudman-etal-2021] - Computed on WikiText-2 token embeddings from GPT (110M), GPT-2 (117M), BERT-base-uncased and DistilBERT-base-uncased. [rudman-etal-2021]

models: BERT-base-uncased · method: IsoScore, PCA
Outlier Dimensions Encode Task-Specific Knowledge (2023)measured

Outlier dimensions persist and can solve GLUE tasks

Details

- Outlier dimensions (variance >=5x the average) that arise in pretraining persist as outliers after fine-tuning regardless of task or seed (GPT-2's exact top-3 persist). [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - A single principal outlier dimension with a learned threshold can complete GLUE binary-classification tasks, with model-dependent accuracy loss: low for GPT-2 (2.85% avg) and ALBERT (6.12%) but high for BERT (16.33%), RoBERTa (19.23%) and DistilBERT (19.12%); Pythia 70M/160M/410M lose 12.58/7.48/11.44%. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - The often-quoted "0-3.4%" figures are QNLI-only (ALBERT 3.42%, Pythia-410M 0.0%), not the cross-task averages; Pythia-410M (11.44% avg) is not low-loss. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - Tested on eight encoder and decoder models (BERT, ALBERT, DistilBERT, RoBERTa, GPT-2, Pythia-70M/160M/410M); the paper names families/sizes, not exact checkpoints. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge]

models: BERT-base-uncased · method: Rogue-dimension analysis
The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: BERT-base-uncased · method: SVD, Intrinsic dimension estimation (TwoNN)
All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality (2021)measured

A few rogue dimensions drive transformer anisotropy

Details

- A handful of "rogue" dimensions, often a single one, with huge off-center means dominate cosine similarity and the anisotropy statistic; the top dimension contributes over 99% of expected cosine in XLNet (0.996 at layer 11; BERT 0.817, GPT-2 0.763, RoBERTa 0.726). [timkey-vanschijndel-2021] - Standardizing or removing ~5 such dimensions restores near-isotropy (A-hat <= 0.25) and improves correlation with human similarity judgments on RG65, WS353, SimLex-999 and SimVerb-3500. [timkey-vanschijndel-2021] - Ablating the rogue dimensions (zeroing their layer-norm gamma and beta) barely changes the output distribution (KL), so they dominate cosine similarity but not model behavior. [timkey-vanschijndel-2021] - Tested on BERT, RoBERTa, GPT-2 and XLNet (base size) plus word2vec (GoogleNews). [timkey-vanschijndel-2021]

models: BERT-base-cased · method: Rogue-dimension analysis, Causal interventions (steering)
Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: BERT-base-uncased, mBERT (BERT-base, Multilingual Cased) · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control
WhiteningBERT: An Easy Unsupervised Sentence Embedding Approach (2021)measured

Whitening removes anisotropy and improves sentence similarity

Details

- Whitening, a parameter-free linear map (mean-center, then rotate and scale by the covariance eigendecomposition), makes sentence embeddings isotropic and improves cosine STS. [huang-etal-2021] - Average STS Spearman rises 62.97 to 67.76 (BERT), 59.53 to 67.72 (RoBERTa) and 64.12 to 68.30 (DistilBERT), but only 71.56 to 71.71 for the already-isotropic LaBSE. [huang-etal-2021] - Averaging the first and last layers (L1+L12) beats any single layer, and token-averaging beats the [CLS] vector by a large margin. [huang-etal-2021] - Main models are BERT-base, RoBERTa-base, DistilBERT and LaBSE, with the effect confirmed across 24 pretrained encoders in the appendix. [huang-etal-2021]

models: BERT-base-uncased · method: Whitening transformation
Whitening Sentence Representations for Better Semantics and Faster Retrieval (2021)measured

Whitening matches BERT-flow at far lower dimension

Details

- Whitening (mean-center plus PCA whitening) makes BERT embeddings isotropic and ranks dimensions by variance, so the whitening map can be truncated to k dimensions with little loss. [su-etal-2021] - Truncated to just k=109 dimensions, BERT-base-whitening (NLI) reaches SICK-R Spearman 66.52, beating full 768-dimensional BERT-flow (NLI) at 65.44 by +1.08. [su-etal-2021] - The advantage over BERT-flow is on SICK-R (not STS-B), and it is achieved at far lower dimensionality. [su-etal-2021] - Tested on BERT-base and BERT-large (uncased). [su-etal-2021]

models: BERT-base-uncased, BERT-large-uncased · method: Whitening transformation

GPT

How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings (2019)measured

Contextualized embeddings occupy a narrow anisotropic cone

Details

- Across BERT, ELMo and GPT-2, contextualized representations occupy a narrow anisotropic cone: two random words have high average cosine similarity, near 1.0 in GPT-2's last layer. [ethayarajh-2019] - Anisotropy increases in upper layers together with context-specificity, and self-similarity falls monotonically with depth in all three models. [ethayarajh-2019] - Fewer than 5% of a word's contextual variance is captured by a single static embedding (anisotropy-adjusted MEV). [ethayarajh-2019] - Tested on BERT-base-cased, the original 2-layer biLSTM ELMo, and GPT-2 small; analysis is observational. [ethayarajh-2019]

models: GPT-2-small · method: Contextuality metrics (self-similarity, intra-sentence similarity, MEV), PCA
Isotropy in the Contextual Embedding Space: Clusters and Manifolds (2021)measured

Clusters and low-dim manifolds explain apparent anisotropy

Details

- Apparent global anisotropy (inter-type cosine ~0.7) is largely an artifact of a few isolated clusters: after k-means clustering and per-cluster mean-shifting, within-cluster cosine isotropy is near zero across every layer. [cai-etal-2021] - Contextual embeddings sit on low intrinsic-dimension manifolds, with Local Intrinsic Dimension 5.6 (BERT), 7.3 (DistilBERT), 6.8 (GPT), 7.0 (GPT-2), 9.1 (ELMo), far below the 768/1024 ambient dimensions. [cai-etal-2021] - GPT and GPT-2 additionally show a Swiss-Roll manifold tied to word frequency. [cai-etal-2021] - Tested on uncased base BERT, DistilBERT, GPT and GPT-2 (Hugging Face) and AllenNLP ELMo (no 5.5B variant); observational. [cai-etal-2021]

models: GPT-1 (OpenAI GPT), GPT-2-small · method: Cluster-corrected isotropy analysis
IsoScore: Measuring the Uniformity of Embedding Space Utilization (2021)measured

IsoScore shows contextual spaces use even fewer dimensions

Details

- IsoScore measures isotropy from the covariance alone (PCA-reorient, length-normalize the variance vector, map to [0,1]), making it mean-agnostic, scale- and rotation-invariant. [rudman-etal-2021] - Every contextualized model scores below 0.18 (GPT and GPT-2 do not isotropically use even one dimension), so they use fewer dimensions than cosine-based estimates implied. [rudman-etal-2021] - IsoScore overturns two prior conclusions: isotropy does not increase in earlier layers, and mean-centering does not recover isotropy (a rotation test holds IsoScore at 0.216 while average cosine swings 0.97-0.99). [rudman-etal-2021] - Computed on WikiText-2 token embeddings from GPT (110M), GPT-2 (117M), BERT-base-uncased and DistilBERT-base-uncased. [rudman-etal-2021]

models: GPT-1 (OpenAI GPT), GPT-2-small · method: IsoScore, PCA
Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding (2025)measured

RoPE massive values concentrate at fixed Q/K dimensions

Details

- In RoPE transformers, "massive values" (per-head L2 norm >5x the mean) concentrate at consistent head-dimension indices in query/key space after RoPE, and are absent in non-RoPE models. [jin-etal-2025-massive-values-self-attention] - Disrupting the massive-value dimensions at prefill collapses contextual-knowledge tasks (Llama-3-8B GSM8K 76.9 to 4.0, IMDB 95.4 to 11.0, Passkey 100 to 9.0) while parametric knowledge survives (Cities 95.0 to 88.2); disrupting non-massive dimensions changes accuracy <+-1%. [jin-etal-2025-massive-values-self-attention] - Quantization that protects large-magnitude channels (AWQ, SmoothQuant) preserves contextual accuracy, whereas GPTQ, which does not, degrades it. [jin-etal-2025-massive-values-self-attention] - Main results on Llama-3-8B-Instruct, Gemma-2-9B-it and Qwen2.5-7B-Instruct, with Llama-2-7B-chat and non-RoPE GPT-2/OPT as contrasts (further appendix models not re-verified here). [jin-etal-2025-massive-values-self-attention]

models: GPT-2-small · method: Geometric analysis, Causal interventions (steering)
The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: GPT-2-small · method: SVD, Intrinsic dimension estimation (TwoNN)
All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality (2021)measured

A few rogue dimensions drive transformer anisotropy

Details

- A handful of "rogue" dimensions, often a single one, with huge off-center means dominate cosine similarity and the anisotropy statistic; the top dimension contributes over 99% of expected cosine in XLNet (0.996 at layer 11; BERT 0.817, GPT-2 0.763, RoBERTa 0.726). [timkey-vanschijndel-2021] - Standardizing or removing ~5 such dimensions restores near-isotropy (A-hat <= 0.25) and improves correlation with human similarity judgments on RG65, WS353, SimLex-999 and SimVerb-3500. [timkey-vanschijndel-2021] - Ablating the rogue dimensions (zeroing their layer-norm gamma and beta) barely changes the output distribution (KL), so they dominate cosine similarity but not model behavior. [timkey-vanschijndel-2021] - Tested on BERT, RoBERTa, GPT-2 and XLNet (base size) plus word2vec (GoogleNews). [timkey-vanschijndel-2021]

models: GPT-2-small · method: Rogue-dimension analysis, Causal interventions (steering)
Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: GPT-2-small · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control

Wan

Steering Video Diffusion Transformers with Massive Activations (2026)measured

Video-DiT massive activations follow a token-position hierarchy

Details

- Video diffusion transformers carry massive-activation channels exceeding 50x the mean activation at a fixed small set of dimensions, consistent across layers and model scales. [cheng-etal-2026-steering-video-diffusion-transformers-massive-activations] - Magnitudes follow a token-position hierarchy: first-latent-frame tokens are largest, with periodic spikes at the VAE's 4x temporal-chunk boundaries, decaying toward the interior over denoising. [cheng-etal-2026-steering-video-diffusion-transformers-massive-activations] - Structured Activation Steering overwrites the top/tail ~8% boundary-token values toward a scaled reference, improving VBench (Wan2.1 81.39 to 81.76, CogVideoX-5B 79.34 to 79.61, Wan2.2 81.74 to 81.93) at negligible cost; a position-agnostic edit instead hurts quality. [cheng-etal-2026-steering-video-diffusion-transformers-massive-activations] - Tested on Wan2.1-T2V-1.3B, Wan2.2-TI2V-5B and CogVideoX-5B (FLUX is an image-DiT baseline only). [cheng-etal-2026-steering-video-diffusion-transformers-massive-activations]

models: Wan2.1-1.3B, Wan2.2-5B · method: Geometric analysis, Activation Steering (Addition)

CogVideoX

Steering Video Diffusion Transformers with Massive Activations (2026)measured

Video-DiT massive activations follow a token-position hierarchy

Details

- Video diffusion transformers carry massive-activation channels exceeding 50x the mean activation at a fixed small set of dimensions, consistent across layers and model scales. [cheng-etal-2026-steering-video-diffusion-transformers-massive-activations] - Magnitudes follow a token-position hierarchy: first-latent-frame tokens are largest, with periodic spikes at the VAE's 4x temporal-chunk boundaries, decaying toward the interior over denoising. [cheng-etal-2026-steering-video-diffusion-transformers-massive-activations] - Structured Activation Steering overwrites the top/tail ~8% boundary-token values toward a scaled reference, improving VBench (Wan2.1 81.39 to 81.76, CogVideoX-5B 79.34 to 79.61, Wan2.2 81.74 to 81.93) at negligible cost; a position-agnostic edit instead hurts quality. [cheng-etal-2026-steering-video-diffusion-transformers-massive-activations] - Tested on Wan2.1-T2V-1.3B, Wan2.2-TI2V-5B and CogVideoX-5B (FLUX is an image-DiT baseline only). [cheng-etal-2026-steering-video-diffusion-transformers-massive-activations]

models: CogVideoX-5B · method: Geometric analysis, Activation Steering (Addition)

CLIP (Contrastive Language-Image Pretraining)

Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning (2022)measured

CLIP's modality gap is a provable, tunable cone effect

Details

- Each modality's embeddings collapse into a narrow cone present even at random initialization; two distinct cones plus low temperature (tau=1/100) leave a persistent image-text modality gap. [liang-etal-2022-modality-gap] - The default gap magnitude ||Delta||=0.82 is the contrastive-loss global minimum, so the gap is a geometric consequence of the cone effect rather than an artifact. [liang-etal-2022-modality-gap] - Causally shifting embeddings along the gap direction changes downstream behavior: widening the gap 0.82 to 0.97 reduces FairFace denigration, with small zero-shot shifts (CIFAR-10 0.9013 to 0.9081). [liang-etal-2022-modality-gap] - Zero-shot accuracy experiments use CLIP ViT-B/16, fairness uses ViT-B/32; the paper also documents the gap in VideoCLIP, ConVIRT and CLASP. [liang-etal-2022-modality-gap]

models: CLIP ViT-B/32, CLIP ViT-B/16 · method: Geometric analysis, Causal interventions (steering)
The Double-Ellipsoid Geometry of CLIP (2025)measured

CLIP embeddings lie on two tilted offset ellipsoids

Details

- Pre-normalization CLIP embeddings for each modality lie on a thin ellipsoidal shell: long-tailed per-feature variance makes it an ellipsoid and off-diagonal covariance tilts it. [levi-gilboa-2025-double-ellipsoid-clip] - The image and text shells are separable and centered well away from the origin (the "double-ellipsoid"); a linear SVM separates the modalities with 100% accuracy using just two features. [levi-gilboa-2025-double-ellipsoid-clip] - The thin-shell approximation is tight (image mu_norm=7.59, 0.18% relative error) and a conformity score tracks generation quality at Pearson 0.9998. [levi-gilboa-2025-double-ellipsoid-clip] - Observational on CLIP ViT-B/32 (primary) and ViT-L/14 (supplementary); interventions are post-hoc embedding shifts. [levi-gilboa-2025-double-ellipsoid-clip]

models: CLIP ViT-B/32, CLIP ViT-L/14 · method: Geometric analysis
It's Not a Modality Gap: Characterizing and Addressing the Contrastive Gap (2024)measured

CLIP's modality gap is a contrastive uniformity failure

Details

- The CLIP "modality gap" is reframed as a contrastive gap: a failure of uniformity in which embeddings occupy a narrow, low-dimensional cone instead of filling the hypersphere. [fahim-etal-2024-contrastive-gap] - Adding Wang-and-Isola uniformity and alignment terms plus a cross-modal uniformity term collapses the centroid gap 0.66 to 0.08 and lowers modality linear separability 1.00 to 0.73. [fahim-etal-2024-contrastive-gap] - The added terms improve multimodal arithmetic (SIMAT 36.02 to 42.18) and zero-shot classification while leaving retrieval roughly flat. [fahim-etal-2024-contrastive-gap] - Tested on CLIP ViT-B/32 fine-tuned on MS-COCO at reduced embedding dimensions (32/64/128D). [fahim-etal-2024-contrastive-gap]

models: CLIP ViT-B/32 · method: Geometric analysis, Training-loss uniformity/alignment regularization
Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders (2026)measured

Same concept, different SAE directions across modalities

Details

- In vision-language models, sparse-autoencoder feature directions for the same concept differ across modalities, keeping substantial angular (cosine-distance) separation even when their activations are highly correlated. [lee-etal-2026-cross-modal-sae-heterogeneity] - The paper proves that a single forced-alignment SAE collapses the two directions to their normalized sum [(phi+psi)/||phi+psi||, 0], sacrificing reconstruction. [lee-etal-2026-cross-modal-sae-heterogeneity] - Modality-specific SAEs plus post-hoc alignment give better cross-modal retrieval and steering (image->text R@1 16.0, text->image R@1 11.4). [lee-etal-2026-cross-modal-sae-heterogeneity] - Directional analysis spans CLIP ViT-B/32, MetaCLIP B/32, OpenCLIP B/32 and SigLIP2. [lee-etal-2026-cross-modal-sae-heterogeneity]

models: CLIP ViT-B/32 · method: TopK (k-sparse) Sparse Autoencoders, Geometric analysis, Causal interventions (steering)
Cross-Modal Redundancy and the Geometry of Vision-Language Embeddings (2026)measured

Iso-energy SAE isolates a bimodal cross-modal subspace

Details

- An iso-energy-regularized aligned SAE splits its dictionary into bimodal atoms (a modality-agnostic shared subspace carrying essentially all cross-modal alignment) and unimodal atoms (per-modality cones that account for the modality gap). [dhimoila-etal-2026-cross-modal-redundancy] - Ablating the unimodal atoms nearly eliminates the modality gap while barely changing retrieval recall (CLIP-B/32 delta_r 0.224 to 0.125 from SAE to SAE-A; functional alignment rho 0.327 to 4.232). [dhimoila-etal-2026-cross-modal-redundancy] - Restricting vector-arithmetic edits to the bimodal subspace keeps them in-distribution and improves OOD retrieval (CLIP 0.97 to 0.77, SigLIP2 0.99 to 0.61); reconstruction R-squared >=0.859. [dhimoila-etal-2026-cross-modal-redundancy] - The paper measures retrieval (delta_r), not post-ablation zero-shot classification; tested on CLIP and OpenCLIP ViT-B/32 and ViT-L/14, SigLIP and SigLIP2. [dhimoila-etal-2026-cross-modal-redundancy]

models: CLIP ViT-B/32, CLIP ViT-L/14 · method: Iso-Energy regularized sparse autoencoder (SAE-A), Causal interventions (steering)
Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE) (2024)measured

Sparse nonnegative concept decomposition enables causal CLIP editing

Details

- CLIP image embeddings decompose as sparse, nonnegative combinations (l0 ~5-20, recommended ~10-30) of a fixed dictionary of ~15k text-encoder concept vectors from LAION captions. [bhalla-etal-2024-splice] - The decomposition nearly preserves zero-shot accuracy (ImageNet 0.552 to 0.552, CIFAR-100 0.750 to 0.768) with cosine reconstruction ~0.75-0.88. [bhalla-etal-2024-splice] - Because embeddings are additive concept mixtures, zeroing a concept edits behavior surgically (removing "glasses" drops a glasses classifier 0.88 to 0.69 while gender stays 0.97 to 0.96). [bhalla-etal-2024-splice] - Tested on OpenCLIP ViT-B/32 (main) and OpenAI CLIP RN50 (appendix). [bhalla-etal-2024-splice]

models: CLIP ResNet-50 · method: Fixed-dictionary sparse recovery, Causal interventions (steering)
Decipher the Modality Gap in Multimodal Contrastive Learning: From Convergent Representations to Pairwise Alignment (2025)measured

Modality gap converges to angle between collapsed hyperplanes

Details

- The modality gap originates from dimension collapse: image and text representations each collapse onto distinct low-dimensional hyperplanes. [yi-etal-2025-decipher-the-modality-gap-in-multimodal-contrastive-learning] - At the contrastive optimum the modality means become orthogonal to the shared subspace, so the gap converges to the smallest angle between the two hyperplanes (Theorem 3). [yi-etal-2025-decipher-the-modality-gap-in-multimodal-contrastive-learning] - On CLIP ViT-B/32 the gap angle is 74.69 deg (CIFAR-10), 74.19 and 71.02 (ImageNet); a shared-subspace projection reduces it to 5.37/30.39/50.40, with an estimated ~212-dim shared space. [yi-etal-2025-decipher-the-modality-gap-in-multimodal-contrastive-learning] - Single model CLIP ViT-B/32; ViT-L/14 and RN50 are cited but not tested, and interventions are post-hoc geometric (projection/translation/rotation). [yi-etal-2025-decipher-the-modality-gap-in-multimodal-contrastive-learning]

models: CLIP ViT-B/32 · method: Geometric analysis, Causal interventions (steering)
On the Modality Gap and the Contrastive Loss in Multi-modal Representation Learning (2026)measured

Low-temperature InfoNCE is the modality gap's cause

Details

- With identical-initialization dual encoders that start perfectly overlapped, the InfoNCE loss actively separates the two modalities at low temperature, a mode failure rather than an initialization cone effect. [mager-etal-2026-on-the-modality-gap-and-the-contrastive-loss-in-multi-modal-representation-learning] - Adding intra-modal negatives (xNCE) prevents this separation across temperatures (MNIST modality-centroid distance 0.75 at tau=0.01 under InfoNCE vs ~0.09-0.13 under xNCE). [mager-etal-2026-on-the-modality-gap-and-the-contrastive-loss-in-multi-modal-representation-learning] - The gap is largely reduced rather than proven eliminated; the mechanism is supported by a theoretical analysis of the InfoNCE temperature term. [mager-etal-2026-on-the-modality-gap-and-the-contrastive-loss-in-multi-modal-representation-learning] - Real-model experiments use CLIP ViT-B/16 on MS-COCO, alongside a custom MNIST vision transformer that carries the mechanistic evidence. [mager-etal-2026-on-the-modality-gap-and-the-contrastive-loss-in-multi-modal-representation-learning]

models: CLIP ViT-B/16 · method: Geometric analysis, Causal interventions (steering), Theoretical / Analytical
Interpreting the Linear Structure of Vision-Language Model Embedding Spaces (2025)measured

VLM SAE concepts are single-modality yet gap-orthogonal

Details

- BatchTopK sparse autoencoders trained on ~600k COCO vision-language embeddings (dictionary 4096) put 99% of the energy in 512 concepts, the other ~3500 sharing 1%. [papadimitriou-etal-2025-vlm-linear-structure] - The energy-dominant concepts are nearly single-modality in activation, yet many of their directions are nearly orthogonal to the modality-gap subspace (accuracy ~0.5 as modality classifiers). [papadimitriou-etal-2025-vlm-linear-structure] - An "SAE projection effect" reconciles this: sparse thresholding interacts with the per-modality input distributions, so a direction can be gap-orthogonal yet fire for one modality; top-512 concepts are seed-stable (0.92) vs 0.16 for low-energy ones. [papadimitriou-etal-2025-vlm-linear-structure] - Observational, on CLIP, SigLIP, SigLIP2 and AIMv2 (the paper states no model variants). [papadimitriou-etal-2025-vlm-linear-structure]

models: CLIP · method: BatchTopK Sparse Autoencoders, Geometric analysis

DistilBERT

Isotropy in the Contextual Embedding Space: Clusters and Manifolds (2021)measured

Clusters and low-dim manifolds explain apparent anisotropy

Details

- Apparent global anisotropy (inter-type cosine ~0.7) is largely an artifact of a few isolated clusters: after k-means clustering and per-cluster mean-shifting, within-cluster cosine isotropy is near zero across every layer. [cai-etal-2021] - Contextual embeddings sit on low intrinsic-dimension manifolds, with Local Intrinsic Dimension 5.6 (BERT), 7.3 (DistilBERT), 6.8 (GPT), 7.0 (GPT-2), 9.1 (ELMo), far below the 768/1024 ambient dimensions. [cai-etal-2021] - GPT and GPT-2 additionally show a Swiss-Roll manifold tied to word frequency. [cai-etal-2021] - Tested on uncased base BERT, DistilBERT, GPT and GPT-2 (Hugging Face) and AllenNLP ELMo (no 5.5B variant); observational. [cai-etal-2021]

models: DistilBERT-base-uncased · method: Cluster-corrected isotropy analysis
IsoScore: Measuring the Uniformity of Embedding Space Utilization (2021)measured

IsoScore shows contextual spaces use even fewer dimensions

Details

- IsoScore measures isotropy from the covariance alone (PCA-reorient, length-normalize the variance vector, map to [0,1]), making it mean-agnostic, scale- and rotation-invariant. [rudman-etal-2021] - Every contextualized model scores below 0.18 (GPT and GPT-2 do not isotropically use even one dimension), so they use fewer dimensions than cosine-based estimates implied. [rudman-etal-2021] - IsoScore overturns two prior conclusions: isotropy does not increase in earlier layers, and mean-centering does not recover isotropy (a rotation test holds IsoScore at 0.216 while average cosine swings 0.97-0.99). [rudman-etal-2021] - Computed on WikiText-2 token embeddings from GPT (110M), GPT-2 (117M), BERT-base-uncased and DistilBERT-base-uncased. [rudman-etal-2021]

models: DistilBERT-base-uncased · method: IsoScore, PCA
Outlier Dimensions Encode Task-Specific Knowledge (2023)measured

Outlier dimensions persist and can solve GLUE tasks

Details

- Outlier dimensions (variance >=5x the average) that arise in pretraining persist as outliers after fine-tuning regardless of task or seed (GPT-2's exact top-3 persist). [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - A single principal outlier dimension with a learned threshold can complete GLUE binary-classification tasks, with model-dependent accuracy loss: low for GPT-2 (2.85% avg) and ALBERT (6.12%) but high for BERT (16.33%), RoBERTa (19.23%) and DistilBERT (19.12%); Pythia 70M/160M/410M lose 12.58/7.48/11.44%. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - The often-quoted "0-3.4%" figures are QNLI-only (ALBERT 3.42%, Pythia-410M 0.0%), not the cross-task averages; Pythia-410M (11.44% avg) is not low-loss. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - Tested on eight encoder and decoder models (BERT, ALBERT, DistilBERT, RoBERTa, GPT-2, Pythia-70M/160M/410M); the paper names families/sizes, not exact checkpoints. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge]

models: DistilBERT-base-uncased · method: Rogue-dimension analysis
WhiteningBERT: An Easy Unsupervised Sentence Embedding Approach (2021)measured

Whitening removes anisotropy and improves sentence similarity

Details

- Whitening, a parameter-free linear map (mean-center, then rotate and scale by the covariance eigendecomposition), makes sentence embeddings isotropic and improves cosine STS. [huang-etal-2021] - Average STS Spearman rises 62.97 to 67.76 (BERT), 59.53 to 67.72 (RoBERTa) and 64.12 to 68.30 (DistilBERT), but only 71.56 to 71.71 for the already-isotropic LaBSE. [huang-etal-2021] - Averaging the first and last layers (L1+L12) beats any single layer, and token-averaging beats the [CLS] vector by a large margin. [huang-etal-2021] - Main models are BERT-base, RoBERTa-base, DistilBERT and LaBSE, with the effect confirmed across 24 pretrained encoders in the appendix. [huang-etal-2021]

models: DistilBERT-base-uncased · method: Whitening transformation

CLAP (Contrastive Language-Audio Pretraining)

COMET: Concept Space Dissection of the Modality Gap in Audio-Text Multimodal Contrastive Embeddings (2026)measured

Audio-text embeddings split into shared head and private tail

Details

- A cross-modal PLS-SVD decomposition splits CLAP audio-text embeddings into a static mean component (the gap), a compact shared semantic head (~100 of 1024 dims carrying most cross-modal variance), and a diffuse modality-private tail. [zhu-etal-2026-comet-audio-modality-gap] - Truncating to the top-100 shared axes (PLSHead) preserves retrieval (Clotho Text->Audio R@1 17.42 to 17.32) while cutting MeanR 42.36 to 36.30. [zhu-etal-2026-comet-audio-modality-gap] - A PCA-head ablation collapses retrieval to R@1 ~0.06/0.00, isolating the PLS shared axes as causally responsible; removing the tail drops captioning BLEU-4 to 5.9-7.1. [zhu-etal-2026-comet-audio-modality-gap] - Tested on CLAP HTSAT-BERT (zero-shot) and the DRCap CLAP backbone. [zhu-etal-2026-comet-audio-modality-gap]

models: CLAP (HTSAT-BERT-ZS, trained on WavCaps), DRCap's CLAP (trained on WavCaps + SoundVECaps) · method: Cross-Modal PLS-SVD Decomposition, Geometric analysis, Causal interventions (steering)

MetaCLIP

Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders (2026)measured

Same concept, different SAE directions across modalities

Details

- In vision-language models, sparse-autoencoder feature directions for the same concept differ across modalities, keeping substantial angular (cosine-distance) separation even when their activations are highly correlated. [lee-etal-2026-cross-modal-sae-heterogeneity] - The paper proves that a single forced-alignment SAE collapses the two directions to their normalized sum [(phi+psi)/||phi+psi||, 0], sacrificing reconstruction. [lee-etal-2026-cross-modal-sae-heterogeneity] - Modality-specific SAEs plus post-hoc alignment give better cross-modal retrieval and steering (image->text R@1 16.0, text->image R@1 11.4). [lee-etal-2026-cross-modal-sae-heterogeneity] - Directional analysis spans CLIP ViT-B/32, MetaCLIP B/32, OpenCLIP B/32 and SigLIP2. [lee-etal-2026-cross-modal-sae-heterogeneity]

models: MetaCLIP ViT-B/32 · method: TopK (k-sparse) Sparse Autoencoders, Geometric analysis, Causal interventions (steering)

OpenCLIP

Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders (2026)measured

Same concept, different SAE directions across modalities

Details

- In vision-language models, sparse-autoencoder feature directions for the same concept differ across modalities, keeping substantial angular (cosine-distance) separation even when their activations are highly correlated. [lee-etal-2026-cross-modal-sae-heterogeneity] - The paper proves that a single forced-alignment SAE collapses the two directions to their normalized sum [(phi+psi)/||phi+psi||, 0], sacrificing reconstruction. [lee-etal-2026-cross-modal-sae-heterogeneity] - Modality-specific SAEs plus post-hoc alignment give better cross-modal retrieval and steering (image->text R@1 16.0, text->image R@1 11.4). [lee-etal-2026-cross-modal-sae-heterogeneity] - Directional analysis spans CLIP ViT-B/32, MetaCLIP B/32, OpenCLIP B/32 and SigLIP2. [lee-etal-2026-cross-modal-sae-heterogeneity]

models: OpenCLIP ViT-B/32 · method: TopK (k-sparse) Sparse Autoencoders, Geometric analysis, Causal interventions (steering)
Cross-Modal Redundancy and the Geometry of Vision-Language Embeddings (2026)measured

Iso-energy SAE isolates a bimodal cross-modal subspace

Details

- An iso-energy-regularized aligned SAE splits its dictionary into bimodal atoms (a modality-agnostic shared subspace carrying essentially all cross-modal alignment) and unimodal atoms (per-modality cones that account for the modality gap). [dhimoila-etal-2026-cross-modal-redundancy] - Ablating the unimodal atoms nearly eliminates the modality gap while barely changing retrieval recall (CLIP-B/32 delta_r 0.224 to 0.125 from SAE to SAE-A; functional alignment rho 0.327 to 4.232). [dhimoila-etal-2026-cross-modal-redundancy] - Restricting vector-arithmetic edits to the bimodal subspace keeps them in-distribution and improves OOD retrieval (CLIP 0.97 to 0.77, SigLIP2 0.99 to 0.61); reconstruction R-squared >=0.859. [dhimoila-etal-2026-cross-modal-redundancy] - The paper measures retrieval (delta_r), not post-ablation zero-shot classification; tested on CLIP and OpenCLIP ViT-B/32 and ViT-L/14, SigLIP and SigLIP2. [dhimoila-etal-2026-cross-modal-redundancy]

models: OpenCLIP ViT-B/32, OpenCLIP ViT-L/14 · method: Iso-Energy regularized sparse autoencoder (SAE-A), Causal interventions (steering)
Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE) (2024)measured

Sparse nonnegative concept decomposition enables causal CLIP editing

Details

- CLIP image embeddings decompose as sparse, nonnegative combinations (l0 ~5-20, recommended ~10-30) of a fixed dictionary of ~15k text-encoder concept vectors from LAION captions. [bhalla-etal-2024-splice] - The decomposition nearly preserves zero-shot accuracy (ImageNet 0.552 to 0.552, CIFAR-100 0.750 to 0.768) with cosine reconstruction ~0.75-0.88. [bhalla-etal-2024-splice] - Because embeddings are additive concept mixtures, zeroing a concept edits behavior surgically (removing "glasses" drops a glasses classifier 0.88 to 0.69 while gender stays 0.97 to 0.96). [bhalla-etal-2024-splice] - Tested on OpenCLIP ViT-B/32 (main) and OpenAI CLIP RN50 (appendix). [bhalla-etal-2024-splice]

models: OpenCLIP ViT-B/32 · method: Fixed-dictionary sparse recovery, Causal interventions (steering)

SigLIP

Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders (2026)measured

Same concept, different SAE directions across modalities

Details

- In vision-language models, sparse-autoencoder feature directions for the same concept differ across modalities, keeping substantial angular (cosine-distance) separation even when their activations are highly correlated. [lee-etal-2026-cross-modal-sae-heterogeneity] - The paper proves that a single forced-alignment SAE collapses the two directions to their normalized sum [(phi+psi)/||phi+psi||, 0], sacrificing reconstruction. [lee-etal-2026-cross-modal-sae-heterogeneity] - Modality-specific SAEs plus post-hoc alignment give better cross-modal retrieval and steering (image->text R@1 16.0, text->image R@1 11.4). [lee-etal-2026-cross-modal-sae-heterogeneity] - Directional analysis spans CLIP ViT-B/32, MetaCLIP B/32, OpenCLIP B/32 and SigLIP2. [lee-etal-2026-cross-modal-sae-heterogeneity]

models: SigLIP 2 · method: TopK (k-sparse) Sparse Autoencoders, Geometric analysis, Causal interventions (steering)
Cross-Modal Redundancy and the Geometry of Vision-Language Embeddings (2026)measured

Iso-energy SAE isolates a bimodal cross-modal subspace

Details

- An iso-energy-regularized aligned SAE splits its dictionary into bimodal atoms (a modality-agnostic shared subspace carrying essentially all cross-modal alignment) and unimodal atoms (per-modality cones that account for the modality gap). [dhimoila-etal-2026-cross-modal-redundancy] - Ablating the unimodal atoms nearly eliminates the modality gap while barely changing retrieval recall (CLIP-B/32 delta_r 0.224 to 0.125 from SAE to SAE-A; functional alignment rho 0.327 to 4.232). [dhimoila-etal-2026-cross-modal-redundancy] - Restricting vector-arithmetic edits to the bimodal subspace keeps them in-distribution and improves OOD retrieval (CLIP 0.97 to 0.77, SigLIP2 0.99 to 0.61); reconstruction R-squared >=0.859. [dhimoila-etal-2026-cross-modal-redundancy] - The paper measures retrieval (delta_r), not post-ablation zero-shot classification; tested on CLIP and OpenCLIP ViT-B/32 and ViT-L/14, SigLIP and SigLIP2. [dhimoila-etal-2026-cross-modal-redundancy]

models: SigLIP, SigLIP 2 · method: Iso-Energy regularized sparse autoencoder (SAE-A), Causal interventions (steering)
Interpreting the Linear Structure of Vision-Language Model Embedding Spaces (2025)measured

VLM SAE concepts are single-modality yet gap-orthogonal

Details

- BatchTopK sparse autoencoders trained on ~600k COCO vision-language embeddings (dictionary 4096) put 99% of the energy in 512 concepts, the other ~3500 sharing 1%. [papadimitriou-etal-2025-vlm-linear-structure] - The energy-dominant concepts are nearly single-modality in activation, yet many of their directions are nearly orthogonal to the modality-gap subspace (accuracy ~0.5 as modality classifiers). [papadimitriou-etal-2025-vlm-linear-structure] - An "SAE projection effect" reconciles this: sparse thresholding interacts with the per-modality input distributions, so a direction can be gap-orthogonal yet fire for one modality; top-512 concepts are seed-stable (0.92) vs 0.16 for low-energy ones. [papadimitriou-etal-2025-vlm-linear-structure] - Observational, on CLIP, SigLIP, SigLIP2 and AIMv2 (the paper states no model variants). [papadimitriou-etal-2025-vlm-linear-structure]

models: SigLIP, SigLIP 2 · method: BatchTopK Sparse Autoencoders, Geometric analysis

LLaVA

What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models (2026)measured

MLLM visual tokens split into sink, dead, alive clusters

Details

- Visual tokens fed into multimodal LLMs partition into three functional clusters, sink (~10%), dead (~30%) and alive (~60%), with near-constant image-agnostic centroids for sink and dead. [fan-etal-2026-visual-tokens-sparsity-redundancy] - Sink-token centroids reach cross-image cosine similarity >0.99 (variance <1e-5) and the largest dead cluster >0.98, a directional collapse of specific functional subgroups. [fan-etal-2026-visual-tokens-sparsity-redundancy] - Pruning all dead tokens raises accuracy +1.0 while removing the same number of random tokens loses -2.6, confirming the typing causally (Table 2, LLaVA-1.5-7B). [fan-etal-2026-visual-tokens-sparsity-redundancy] - Primary model LLaVA-1.5-7B, with generalization to LLaVA-1.5-13B, InternVL-3-8B and Qwen2.5-VL. [fan-etal-2026-visual-tokens-sparsity-redundancy]

models: LLaVA-1.5-7B, LLaVA-1.5-13B · method: Activation-cluster functional typing, Causal interventions (steering)
Beyond Semantics: Rediscovering Spatial Awareness in Vision-Language Models (2025)measured

Oversized vision-token norms suppress VLM spatial awareness

Details

- In LLaVA-1.5-7B, vision-token L2 norms are 1-2 orders of magnitude (up to 3) larger than text-token norms, so vision tokens dominate attention logits. [qi-etal-2025-beyond-semantics-spatial-awareness-vlms] - These oversized norms suppress positional sensitivity, making the model behave like a bag of tokens (shown by the Token Permutation Test and Spatial Compression Study). [qi-etal-2025-beyond-semantics-spatial-awareness-vlms] - RMS-normalizing vision embeddings toward the text-norm scale restores spatial awareness (+2.67 on the 2D spatial task from norm alone, +8.17 with multilayer features; VQAv2 +0.97, CV-Bench-2D +3.32). [qi-etal-2025-beyond-semantics-spatial-awareness-vlms] - Diagnostics are the Token Permutation Test, Spatial Compression Study, Embedding Norm Analysis, Attention Entropy and Attention Visualization. [qi-etal-2025-beyond-semantics-spatial-awareness-vlms]

models: LLaVA-1.5-7B · method: Geometric analysis, Causal interventions (steering)

InternVL

What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models (2026)measured

MLLM visual tokens split into sink, dead, alive clusters

Details

- Visual tokens fed into multimodal LLMs partition into three functional clusters, sink (~10%), dead (~30%) and alive (~60%), with near-constant image-agnostic centroids for sink and dead. [fan-etal-2026-visual-tokens-sparsity-redundancy] - Sink-token centroids reach cross-image cosine similarity >0.99 (variance <1e-5) and the largest dead cluster >0.98, a directional collapse of specific functional subgroups. [fan-etal-2026-visual-tokens-sparsity-redundancy] - Pruning all dead tokens raises accuracy +1.0 while removing the same number of random tokens loses -2.6, confirming the typing causally (Table 2, LLaVA-1.5-7B). [fan-etal-2026-visual-tokens-sparsity-redundancy] - Primary model LLaVA-1.5-7B, with generalization to LLaVA-1.5-13B, InternVL-3-8B and Qwen2.5-VL. [fan-etal-2026-visual-tokens-sparsity-redundancy]

models: InternVL3-8B · method: Activation-cluster functional typing, Causal interventions (steering)

Gao et al. (2019) tied-embedding NLG models

Representation Degeneration Problem in Training Natural Language Generation Models (2019)measured

Weight-tied embeddings degenerate into a narrow cone

Details

- In real trained NLG models with tied input-output embeddings, the word vectors degenerate into a narrow cone with nearly all pairwise cosine similarities positive. [gao-etal-2019-representation-degeneration-problem] - The cause is MLE under weight tying: non-target (especially rare and unseen) word embeddings are pushed toward a shared negative direction of the hidden states, formally when the convex hull of hidden states excludes the origin (made likely by layer normalization). [gao-etal-2019-representation-degeneration-problem] - A cosine-similarity regularizer, MLE-CosReg (gamma=1, penalizing mean pairwise cosine, not "MMI"), flattens the singular-value spectrum and improves WMT'14 BLEU (+1.08 En-De, +0.93 De-En) and WikiText-2 perplexity (52.0 to 50.0). [gao-etal-2019-representation-degeneration-problem] - Studied on Transformer base/big NMT (WMT'14 En-De/De-En) and an AWD-LSTM LM (WikiText-2). [gao-etal-2019-representation-degeneration-problem]

models: Tied-input-output-embedding NMT + LM (WMT'14 / WikiText-2) · method: SVD

Llama

Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding (2025)measured

RoPE massive values concentrate at fixed Q/K dimensions

Details

- In RoPE transformers, "massive values" (per-head L2 norm >5x the mean) concentrate at consistent head-dimension indices in query/key space after RoPE, and are absent in non-RoPE models. [jin-etal-2025-massive-values-self-attention] - Disrupting the massive-value dimensions at prefill collapses contextual-knowledge tasks (Llama-3-8B GSM8K 76.9 to 4.0, IMDB 95.4 to 11.0, Passkey 100 to 9.0) while parametric knowledge survives (Cities 95.0 to 88.2); disrupting non-massive dimensions changes accuracy <+-1%. [jin-etal-2025-massive-values-self-attention] - Quantization that protects large-magnitude channels (AWQ, SmoothQuant) preserves contextual accuracy, whereas GPTQ, which does not, degrades it. [jin-etal-2025-massive-values-self-attention] - Main results on Llama-3-8B-Instruct, Gemma-2-9B-it and Qwen2.5-7B-Instruct, with Llama-2-7B-chat and non-RoPE GPT-2/OPT as contrasts (further appendix models not re-verified here). [jin-etal-2025-massive-values-self-attention]

models: Llama-3-8B-Instruct, Llama-2-7B-Chat · method: Geometric analysis, Causal interventions (steering)
Analysing the Residual Stream of Language Models Under Knowledge Conflicts (2024)measured

Residual-stream probes detect knowledge conflict and reliance

Details

- Logistic-regression probes on the mid-layer residual stream detect parametric-vs-contextual knowledge conflict with ~90% accuracy (peaking around layer 14 of Llama-3-8B on NQSwap). [zhao-etal-2024-analysing-the-residual-stream-of-language-models-under-knowledge-conflicts] - The same probes predict which knowledge source the model will rely on before generation (peaking around layers 16-17). [zhao-etal-2024-analysing-the-residual-stream-of-language-models-under-knowledge-conflicts] - When the model relies on contextual knowledge the residual stream is distinctly more skewed than for parametric knowledge (most pronounced in layers ~20-30). [zhao-etal-2024-analysing-the-residual-stream-of-language-models-under-knowledge-conflicts] - Tested on base Llama-3-8B and Llama-2-7B over NQSwap, Macnoise and ConflictQA; observational. [zhao-etal-2024-analysing-the-residual-stream-of-language-models-under-knowledge-conflicts]

models: Llama-3-8B, Llama-2-7B · method: Linear probing
Learning Uncertainty from Sequential Internal Dispersion in Large Language Models (2026)measured

Cross-layer dispersion of hidden states detects hallucination

Details

- A per-token signature combines the log-pseudo-determinant of the regularized cross-layer hidden-state covariance (a generalized variance) with the circular variance of normalized per-layer hidden states and predictive entropy. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Hallucination-prone tokens show higher cross-layer dispersion, and the signature beats the strongest baseline (Ministral-8B average AUC 78.11 vs 75.24, FPR@95 -7.53). [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Cross-layer (all-pairs) dispersion captures uncertainty that successive-layer methods miss and generalizes better out-of-distribution; it does not compute intrinsic dimension. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Tested on Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct and Ministral-8B-Instruct, with Qwen-3-4B and Qwen-3-14B in the appendix. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models]

models: Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct · method: Geometric analysis
The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: Llama-2-7B, Llama-2-7B-Chat · method: SVD, Intrinsic dimension estimation (TwoNN)

Gemma

Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding (2025)measured

RoPE massive values concentrate at fixed Q/K dimensions

Details

- In RoPE transformers, "massive values" (per-head L2 norm >5x the mean) concentrate at consistent head-dimension indices in query/key space after RoPE, and are absent in non-RoPE models. [jin-etal-2025-massive-values-self-attention] - Disrupting the massive-value dimensions at prefill collapses contextual-knowledge tasks (Llama-3-8B GSM8K 76.9 to 4.0, IMDB 95.4 to 11.0, Passkey 100 to 9.0) while parametric knowledge survives (Cities 95.0 to 88.2); disrupting non-massive dimensions changes accuracy <+-1%. [jin-etal-2025-massive-values-self-attention] - Quantization that protects large-magnitude channels (AWQ, SmoothQuant) preserves contextual accuracy, whereas GPTQ, which does not, degrades it. [jin-etal-2025-massive-values-self-attention] - Main results on Llama-3-8B-Instruct, Gemma-2-9B-it and Qwen2.5-7B-Instruct, with Llama-2-7B-chat and non-RoPE GPT-2/OPT as contrasts (further appendix models not re-verified here). [jin-etal-2025-massive-values-self-attention]

models: Gemma-2-9B-it · method: Geometric analysis, Causal interventions (steering)

Qwen

Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding (2025)measured

RoPE massive values concentrate at fixed Q/K dimensions

Details

- In RoPE transformers, "massive values" (per-head L2 norm >5x the mean) concentrate at consistent head-dimension indices in query/key space after RoPE, and are absent in non-RoPE models. [jin-etal-2025-massive-values-self-attention] - Disrupting the massive-value dimensions at prefill collapses contextual-knowledge tasks (Llama-3-8B GSM8K 76.9 to 4.0, IMDB 95.4 to 11.0, Passkey 100 to 9.0) while parametric knowledge survives (Cities 95.0 to 88.2); disrupting non-massive dimensions changes accuracy <+-1%. [jin-etal-2025-massive-values-self-attention] - Quantization that protects large-magnitude channels (AWQ, SmoothQuant) preserves contextual accuracy, whereas GPTQ, which does not, degrades it. [jin-etal-2025-massive-values-self-attention] - Main results on Llama-3-8B-Instruct, Gemma-2-9B-it and Qwen2.5-7B-Instruct, with Llama-2-7B-chat and non-RoPE GPT-2/OPT as contrasts (further appendix models not re-verified here). [jin-etal-2025-massive-values-self-attention]

models: Qwen2.5-7B-Instruct · method: Geometric analysis, Causal interventions (steering)
Learning Uncertainty from Sequential Internal Dispersion in Large Language Models (2026)measured

Cross-layer dispersion of hidden states detects hallucination

Details

- A per-token signature combines the log-pseudo-determinant of the regularized cross-layer hidden-state covariance (a generalized variance) with the circular variance of normalized per-layer hidden states and predictive entropy. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Hallucination-prone tokens show higher cross-layer dispersion, and the signature beats the strongest baseline (Ministral-8B average AUC 78.11 vs 75.24, FPR@95 -7.53). [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Cross-layer (all-pairs) dispersion captures uncertainty that successive-layer methods miss and generalizes better out-of-distribution; it does not compute intrinsic dimension. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Tested on Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct and Ministral-8B-Instruct, with Qwen-3-4B and Qwen-3-14B in the appendix. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models]

models: Qwen3-4B, Qwen3-14B · method: Geometric analysis
Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: Qwen2.5-1.5B, Qwen2.5-7B · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control
The Myth of Expert Specialization in MoEs: Why Routing Reflects Geometry, Not Necessarily Domain Expertise (2026)measured

MoE routing reflects hidden-state geometry, not domain expertise

Details

- Under the load-balancing auxiliary loss, MoE routers are pushed to suppress the shared dominant direction of their hidden states (Proposition 2), so routing tracks hidden-state geometry rather than domain content. [wang-etal-2026-myth-of-expert-specialization] - Hidden-state cosine similarity predicts expert-usage similarity, and two different models solving the same math problem share ~60% of their most-used experts, the same overlap as one model across different questions. [wang-etal-2026-myth-of-expert-specialization] - Prefill expert usage is nearly identical across semantically different inputs and diverges only during generation, with deep-layer router collapse. [wang-etal-2026-myth-of-expert-specialization] - Studied on five production MoE models: gpt-oss-20B, ERNIE-4.5-21B-A3B-Base, Qwen-3-30B, Ling-mini and Trinity-Mini-Base (Moonlight-16B is a comparison only); observational. [wang-etal-2026-myth-of-expert-specialization]

models: Qwen3-30B-A3B-Base · method: SVD, Geometric analysis

OPT

Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding (2025)measured

RoPE massive values concentrate at fixed Q/K dimensions

Details

- In RoPE transformers, "massive values" (per-head L2 norm >5x the mean) concentrate at consistent head-dimension indices in query/key space after RoPE, and are absent in non-RoPE models. [jin-etal-2025-massive-values-self-attention] - Disrupting the massive-value dimensions at prefill collapses contextual-knowledge tasks (Llama-3-8B GSM8K 76.9 to 4.0, IMDB 95.4 to 11.0, Passkey 100 to 9.0) while parametric knowledge survives (Cities 95.0 to 88.2); disrupting non-massive dimensions changes accuracy <+-1%. [jin-etal-2025-massive-values-self-attention] - Quantization that protects large-magnitude channels (AWQ, SmoothQuant) preserves contextual accuracy, whereas GPTQ, which does not, degrades it. [jin-etal-2025-massive-values-self-attention] - Main results on Llama-3-8B-Instruct, Gemma-2-9B-it and Qwen2.5-7B-Instruct, with Llama-2-7B-chat and non-RoPE GPT-2/OPT as contrasts (further appendix models not re-verified here). [jin-etal-2025-massive-values-self-attention]

models: OPT-350M, OPT-2.7B · method: Geometric analysis, Causal interventions (steering)
The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: OPT-13B · method: SVD, Intrinsic dimension estimation (TwoNN)

Lanza et al. (2026) Continual-Learning ResNet-18 Feature Extractors

Degradation of Feature Space in Continual Learning (2026)measured

Real ResNet-18 feature spaces lose isotropy under continual learning, and forcing it back up hurts accuracy

Details

Lanza, Pereira, Miozzo, Angelats & Dini measure IsoScore and a generalized IsoEntropy (covariance-eigenvalue-spectrum uniformity metrics) on penultimate-layer features of real ResNet-18 encoders trained via SupCon, Co2L, and two novel neural-collapse-inspired variants (SupCP, NCI) on CIFAR-10 and CIFAR-100 under three class-incremental continual-learning splits (50+50, 40+30+30, 20x5), comparing against non-continual centralized training [lanza-etal-2026-degradation-feature-space-continual-learning] IsoScore and IsoEntropy decline as more incremental experiences accumulate for most methods (SupCon accuracy drops from 94.98% centralized to 47.58% under the 20x5 CIFAR-10 split), but isotropy and accuracy are not consistently correlated across methods -- e.g. SupCP shows higher isotropy than NCI on CIFAR-100 while achieving lower accuracy [lanza-etal-2026-degradation-feature-space-continual-learning] Adding a differentiable isotropy regularizer to the Co2L loss causally raises IsoScore (from 0.023 to 0.969 at regularization strength 0.5 on CIFAR-10 20x5) but collapses accuracy from 70.64% to 14.74%, showing that forcibly restoring isotropy is actively harmful rather than protective in this continual-learning setting [lanza-etal-2026-degradation-feature-space-continual-learning]

models: ResNet-18 + MLP projector (CIFAR-10, class-incremental continual learning), ResNet-18 + MLP projector (CIFAR-100, class-incremental continual learning) · method:

Mannes (2026) Positional-Encoding ViT-S Ablation Suite

Positional Encodings Anchor Spatial Structure in Vision Transformers: A Geometric Perspective on Robustness (2026)measured

Positional encodings anchor index-based spatial structure in ViTs

Details

- In a ViT-S/16 trained from scratch on ImageNet-100, no-PE models build content-driven spatial structure, whereas every positional-encoding scheme shifts representations to an index-anchored organization stable under token permutation (SSDC ~0 for no-PE vs 0.3-0.47 for PE). [mannes-2026-positional-encodings-anchor-spatial-structure-vits] - Robustness tracks the integrity of this positional reference frame, not the encoding type: JPEG-q5 fragility is ~0.30 (RoPE), ~0.43 (absolute/sinusoidal) and ~0.63-0.66 (no-PE or permutation-trained). [mannes-2026-positional-encodings-anchor-spatial-structure-vits] - Scaling positional-embedding magnitude down removes the index anchoring and raises fragility in lockstep; the authors frame this as intervention-based evidence, not a fully isolated causal mechanism. [mannes-2026-positional-encodings-anchor-spatial-structure-vits] - Single model: ViT-S/16 (~22M params) with absolute, sinusoidal, rotary and no-PE variants. [mannes-2026-positional-encodings-anchor-spatial-structure-vits]

models: ViT-S/16 (12 layers, 8 heads, dim 384, trained from scratch on ImageNet-100) · method: Geometric analysis, Causal interventions (steering)

word2vec

All-but-the-Top: Simple and Effective Postprocessing for Word Representations (2018)measured

Word embeddings share a mean plus top-PC anisotropy

Details

- Word embeddings share a large nonzero common mean vector plus energy concentrated in ~8-10 dominant directions that encode frequency, making them strongly anisotropic. [mu-viswanath-2018-all-but-the-top] - The all-but-the-top postprocessing (subtract the mean, project out the top D principal components; D=3 for word2vec, D=2 for GloVe, rule D~=d/100) makes vectors more isotropic. [mu-viswanath-2018-all-but-the-top] - The correction causally improves downstream tasks with no retraining: word similarity +1.7% average (GloVe Rare-Words +5.63%), STS +4%, and text classification in 34 of 40 settings (+2.85%). [mu-viswanath-2018-all-but-the-top] - Measured on word2vec (GoogleNews) and GloVe (Common Crawl, 2.2M-vocab / 840B). [mu-viswanath-2018-all-but-the-top]

models: word2vec (Google News, 300d) · method: PCA, Causal interventions (steering)
All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality (2021)measured

A few rogue dimensions drive transformer anisotropy

Details

- A handful of "rogue" dimensions, often a single one, with huge off-center means dominate cosine similarity and the anisotropy statistic; the top dimension contributes over 99% of expected cosine in XLNet (0.996 at layer 11; BERT 0.817, GPT-2 0.763, RoBERTa 0.726). [timkey-vanschijndel-2021] - Standardizing or removing ~5 such dimensions restores near-isotropy (A-hat <= 0.25) and improves correlation with human similarity judgments on RG65, WS353, SimLex-999 and SimVerb-3500. [timkey-vanschijndel-2021] - Ablating the rogue dimensions (zeroing their layer-norm gamma and beta) barely changes the output distribution (KL), so they dominate cosine similarity but not model behavior. [timkey-vanschijndel-2021] - Tested on BERT, RoBERTa, GPT-2 and XLNet (base size) plus word2vec (GoogleNews). [timkey-vanschijndel-2021]

models: word2vec (Google News, 300d) · method: Rogue-dimension analysis, Causal interventions (steering)

GloVe

All-but-the-Top: Simple and Effective Postprocessing for Word Representations (2018)measured

Word embeddings share a mean plus top-PC anisotropy

Details

- Word embeddings share a large nonzero common mean vector plus energy concentrated in ~8-10 dominant directions that encode frequency, making them strongly anisotropic. [mu-viswanath-2018-all-but-the-top] - The all-but-the-top postprocessing (subtract the mean, project out the top D principal components; D=3 for word2vec, D=2 for GloVe, rule D~=d/100) makes vectors more isotropic. [mu-viswanath-2018-all-but-the-top] - The correction causally improves downstream tasks with no retraining: word similarity +1.7% average (GloVe Rare-Words +5.63%), STS +4%, and text classification in 34 of 40 settings (+2.85%). [mu-viswanath-2018-all-but-the-top] - Measured on word2vec (GoogleNews) and GloVe (Common Crawl, 2.2M-vocab / 840B). [mu-viswanath-2018-all-but-the-top]

models: GloVe (840B token Common Crawl) · method: PCA, Causal interventions (steering)

Mistral

Learning Uncertainty from Sequential Internal Dispersion in Large Language Models (2026)measured

Cross-layer dispersion of hidden states detects hallucination

Details

- A per-token signature combines the log-pseudo-determinant of the regularized cross-layer hidden-state covariance (a generalized variance) with the circular variance of normalized per-layer hidden states and predictive entropy. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Hallucination-prone tokens show higher cross-layer dispersion, and the signature beats the strongest baseline (Ministral-8B average AUC 78.11 vs 75.24, FPR@95 -7.53). [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Cross-layer (all-pairs) dispersion captures uncertainty that successive-layer methods miss and generalizes better out-of-distribution; it does not compute intrinsic dimension. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models] - Tested on Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct and Ministral-8B-Instruct, with Qwen-3-4B and Qwen-3-14B in the appendix. [srey-etal-2026-learning-uncertainty-from-sequential-internal-dispersion-in-large-language-models]

models: Ministral-8B-Instruct-2410 · method: Geometric analysis
Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: SFR-Embedding-Mistral, Mistral-7B · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control

ALBERT

Outlier Dimensions Encode Task-Specific Knowledge (2023)measured

Outlier dimensions persist and can solve GLUE tasks

Details

- Outlier dimensions (variance >=5x the average) that arise in pretraining persist as outliers after fine-tuning regardless of task or seed (GPT-2's exact top-3 persist). [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - A single principal outlier dimension with a learned threshold can complete GLUE binary-classification tasks, with model-dependent accuracy loss: low for GPT-2 (2.85% avg) and ALBERT (6.12%) but high for BERT (16.33%), RoBERTa (19.23%) and DistilBERT (19.12%); Pythia 70M/160M/410M lose 12.58/7.48/11.44%. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - The often-quoted "0-3.4%" figures are QNLI-only (ALBERT 3.42%, Pythia-410M 0.0%), not the cross-task averages; Pythia-410M (11.44% avg) is not low-loss. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - Tested on eight encoder and decoder models (BERT, ALBERT, DistilBERT, RoBERTa, GPT-2, Pythia-70M/160M/410M); the paper names families/sizes, not exact checkpoints. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge]

models: ALBERT-base-v1 · method: Rogue-dimension analysis
The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: ALBERT-base-v1 · method: SVD, Intrinsic dimension estimation (TwoNN)

RoBERTa

Outlier Dimensions Encode Task-Specific Knowledge (2023)measured

Outlier dimensions persist and can solve GLUE tasks

Details

- Outlier dimensions (variance >=5x the average) that arise in pretraining persist as outliers after fine-tuning regardless of task or seed (GPT-2's exact top-3 persist). [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - A single principal outlier dimension with a learned threshold can complete GLUE binary-classification tasks, with model-dependent accuracy loss: low for GPT-2 (2.85% avg) and ALBERT (6.12%) but high for BERT (16.33%), RoBERTa (19.23%) and DistilBERT (19.12%); Pythia 70M/160M/410M lose 12.58/7.48/11.44%. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - The often-quoted "0-3.4%" figures are QNLI-only (ALBERT 3.42%, Pythia-410M 0.0%), not the cross-task averages; Pythia-410M (11.44% avg) is not low-loss. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - Tested on eight encoder and decoder models (BERT, ALBERT, DistilBERT, RoBERTa, GPT-2, Pythia-70M/160M/410M); the paper names families/sizes, not exact checkpoints. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge]

models: RoBERTa-base · method: Rogue-dimension analysis
The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: RoBERTa-base · method: SVD, Intrinsic dimension estimation (TwoNN)
All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality (2021)measured

A few rogue dimensions drive transformer anisotropy

Details

- A handful of "rogue" dimensions, often a single one, with huge off-center means dominate cosine similarity and the anisotropy statistic; the top dimension contributes over 99% of expected cosine in XLNet (0.996 at layer 11; BERT 0.817, GPT-2 0.763, RoBERTa 0.726). [timkey-vanschijndel-2021] - Standardizing or removing ~5 such dimensions restores near-isotropy (A-hat <= 0.25) and improves correlation with human similarity judgments on RG65, WS353, SimLex-999 and SimVerb-3500. [timkey-vanschijndel-2021] - Ablating the rogue dimensions (zeroing their layer-norm gamma and beta) barely changes the output distribution (KL), so they dominate cosine similarity but not model behavior. [timkey-vanschijndel-2021] - Tested on BERT, RoBERTa, GPT-2 and XLNet (base size) plus word2vec (GoogleNews). [timkey-vanschijndel-2021]

models: RoBERTa-base · method: Rogue-dimension analysis, Causal interventions (steering)
Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: RoBERTa-base · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control
Lost in State Space: Probing Frozen Mamba Representations (2026)measured

Frozen Mamba collapses into an extreme anisotropy cone

Details

- Frozen Mamba-130m activations collapse into an extremely narrow anisotropy cone: mean pairwise cosine similarity 0.9999 (std 4.4e-5), against RoBERTa-base's ~0.9839. [wagh-singh-2026-lost-in-state-space-probing-frozen-mamba] - The collapse cripples frozen probes (CoLA MCC 0.000 for the raw final state; patched-Mamba 0.246 vs RoBERTa 0.475; STS-B 0.316 vs 0.650). [wagh-singh-2026-lost-in-state-space-probing-frozen-mamba] - A geometry-targeted orthogonal write-injection (eta=0.5) fails to widen the cone and worsens STS-B (0.316 to 0.259), a candid negative result implying geometry correction needs training-time adaptation. [wagh-singh-2026-lost-in-state-space-probing-frozen-mamba] - The comparison model is RoBERTa-base, not BERT, and the "~10x-tighter" figure is the angular ratio between cosine 0.9999 and 0.99, not a measured Mamba-vs-baseline cone ratio. [wagh-singh-2026-lost-in-state-space-probing-frozen-mamba]

models: RoBERTa-base · method: Linear probing, Causal interventions (steering)
WhiteningBERT: An Easy Unsupervised Sentence Embedding Approach (2021)measured

Whitening removes anisotropy and improves sentence similarity

Details

- Whitening, a parameter-free linear map (mean-center, then rotate and scale by the covariance eigendecomposition), makes sentence embeddings isotropic and improves cosine STS. [huang-etal-2021] - Average STS Spearman rises 62.97 to 67.76 (BERT), 59.53 to 67.72 (RoBERTa) and 64.12 to 68.30 (DistilBERT), but only 71.56 to 71.71 for the already-isotropic LaBSE. [huang-etal-2021] - Averaging the first and last layers (L1+L12) beats any single layer, and token-averaging beats the [CLS] vector by a large margin. [huang-etal-2021] - Main models are BERT-base, RoBERTa-base, DistilBERT and LaBSE, with the effect confirmed across 24 pretrained encoders in the appendix. [huang-etal-2021]

models: RoBERTa-base · method: Whitening transformation

GPT-2

Outlier Dimensions Encode Task-Specific Knowledge (2023)measured

Outlier dimensions persist and can solve GLUE tasks

Details

- Outlier dimensions (variance >=5x the average) that arise in pretraining persist as outliers after fine-tuning regardless of task or seed (GPT-2's exact top-3 persist). [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - A single principal outlier dimension with a learned threshold can complete GLUE binary-classification tasks, with model-dependent accuracy loss: low for GPT-2 (2.85% avg) and ALBERT (6.12%) but high for BERT (16.33%), RoBERTa (19.23%) and DistilBERT (19.12%); Pythia 70M/160M/410M lose 12.58/7.48/11.44%. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - The often-quoted "0-3.4%" figures are QNLI-only (ALBERT 3.42%, Pythia-410M 0.0%), not the cross-task averages; Pythia-410M (11.44% avg) is not low-loss. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - Tested on eight encoder and decoder models (BERT, ALBERT, DistilBERT, RoBERTa, GPT-2, Pythia-70M/160M/410M); the paper names families/sizes, not exact checkpoints. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge]

models: GPT-2 Small · method: Rogue-dimension analysis

Pythia

Outlier Dimensions Encode Task-Specific Knowledge (2023)measured

Outlier dimensions persist and can solve GLUE tasks

Details

- Outlier dimensions (variance >=5x the average) that arise in pretraining persist as outliers after fine-tuning regardless of task or seed (GPT-2's exact top-3 persist). [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - A single principal outlier dimension with a learned threshold can complete GLUE binary-classification tasks, with model-dependent accuracy loss: low for GPT-2 (2.85% avg) and ALBERT (6.12%) but high for BERT (16.33%), RoBERTa (19.23%) and DistilBERT (19.12%); Pythia 70M/160M/410M lose 12.58/7.48/11.44%. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - The often-quoted "0-3.4%" figures are QNLI-only (ALBERT 3.42%, Pythia-410M 0.0%), not the cross-task averages; Pythia-410M (11.44% avg) is not low-loss. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge] - Tested on eight encoder and decoder models (BERT, ALBERT, DistilBERT, RoBERTa, GPT-2, Pythia-70M/160M/410M); the paper names families/sizes, not exact checkpoints. [rudman-etal-2023-outlier-dimensions-encode-task-specific-knowledge]

models: Pythia-70M, Pythia-160M, Pythia-410M · method: Rogue-dimension analysis
The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: Pythia-2.8B · method: SVD, Intrinsic dimension estimation (TwoNN)
Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: Pythia-410M · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control

GPT-J

The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: GPT-J-6B · method: SVD, Intrinsic dimension estimation (TwoNN)

BLOOM

The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: BLOOM-560M, BLOOM-3B · method: SVD, Intrinsic dimension estimation (TwoNN)

Falcon

The Shape of Learning: Anisotropy and Intrinsic Dimensions in Transformer-Based Models (2024)measured

Decoder-only models peak anisotropy in middle layers

Details

- Decoder-only models develop a bell-shaped anisotropy profile peaking in the middle layers, unlike encoders' flatter profile, measured by an SVD anisotropy score sigma_1^2/sum(sigma_i^2). [razzhigaev-etal-2024-shape-of-learning] - Intrinsic dimension (TwoNN) follows a two-phase pattern of early growth then decline, and decoder anisotropy rises through pretraining before stabilizing. [razzhigaev-etal-2024-shape-of-learning] - Spans encoders (BERT, RoBERTa, ALBERT) and decoders (OPT, Llama-2 and Llama-2-chat, GPT-2, GPT-J, Falcon-7B and Falcon-7B-Instruct, BLOOM, Pythia); observational on enwik8 character-level embeddings. [razzhigaev-etal-2024-shape-of-learning]

models: Falcon-7B, Falcon-7B-Instruct · method: SVD, Intrinsic dimension estimation (TwoNN)

XLNet

All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality (2021)measured

A few rogue dimensions drive transformer anisotropy

Details

- A handful of "rogue" dimensions, often a single one, with huge off-center means dominate cosine similarity and the anisotropy statistic; the top dimension contributes over 99% of expected cosine in XLNet (0.996 at layer 11; BERT 0.817, GPT-2 0.763, RoBERTa 0.726). [timkey-vanschijndel-2021] - Standardizing or removing ~5 such dimensions restores near-isotropy (A-hat <= 0.25) and improves correlation with human similarity judgments on RG65, WS353, SimLex-999 and SimVerb-3500. [timkey-vanschijndel-2021] - Ablating the rogue dimensions (zeroing their layer-norm gamma and beta) barely changes the output distribution (KL), so they dominate cosine similarity but not model behavior. [timkey-vanschijndel-2021] - Tested on BERT, RoBERTa, GPT-2 and XLNet (base size) plus word2vec (GoogleNews). [timkey-vanschijndel-2021]

models: XLNet-base-cased · method: Rogue-dimension analysis, Causal interventions (steering)

MiniLM

Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: all-MiniLM-L6-v2, all-MiniLM-L12-v2 · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control

MPNet

Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: all-mpnet-base-v2, paraphrase-mpnet-base-v2 · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control

BGE

Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: BGE-base-en-v1.5, BGE-large-en-v1.5 · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control

E5

Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: E5-large-v2, multilingual-e5-large, e5-mistral-7b-instruct · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control

ELECTRA

Anisotropy Decides Cosine vs. Rank Metrics for Text Embeddings (2026)measured

Dominant-dimension variance decides cosine vs rank metrics

Details

- Across 19 text encoders, the variance share held by the single dominant ("rogue") dimension predicts whether rank/L1-type metrics or cosine win (rank correlation 0.86, linear 0.95). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - A dominance threshold ~0.01 splits crowded (9) from well-spread (10) encoders; on crowded encoders the best alternative beats cosine by 0.055 Spearman on average (87% of cells) versus 0.001 for well-spread. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Projecting out the top principal directions (all-but-the-top) erases 87% of the advantage on crowded encoders versus ~10% for random directions, so geometry, not the training method, decides the optimal metric. [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings] - Evaluated on 10 contrastive embedders (MiniLM-L6/L12, MPNet, paraphrase-MPNet, BGE-base/large, E5-large, multilingual-E5-large, E5-Mistral-7B, SFR-Embedding-Mistral) and 9 base LMs (BERT, RoBERTa, ELECTRA, mBERT, GPT-2, Pythia-410M, Qwen2.5-1.5B/7B, Mistral-7B). [parupudi-2026-anisotropy-decides-cosine-vs-rank-metrics-for-text-embeddings]

models: ELECTRA-base · method: IsoScore, Rogue-dimension analysis, Geometric analysis, PCA-derived semantic-subspace ablation, validated against a randomized-subspace control

AIMv2

Interpreting the Linear Structure of Vision-Language Model Embedding Spaces (2025)measured

VLM SAE concepts are single-modality yet gap-orthogonal

Details

- BatchTopK sparse autoencoders trained on ~600k COCO vision-language embeddings (dictionary 4096) put 99% of the energy in 512 concepts, the other ~3500 sharing 1%. [papadimitriou-etal-2025-vlm-linear-structure] - The energy-dominant concepts are nearly single-modality in activation, yet many of their directions are nearly orthogonal to the modality-gap subspace (accuracy ~0.5 as modality classifiers). [papadimitriou-etal-2025-vlm-linear-structure] - An "SAE projection effect" reconciles this: sparse thresholding interacts with the per-modality input distributions, so a direction can be gap-orthogonal yet fire for one modality; top-512 concepts are seed-stable (0.92) vs 0.16 for low-energy ones. [papadimitriou-etal-2025-vlm-linear-structure] - Observational, on CLIP, SigLIP, SigLIP2 and AIMv2 (the paper states no model variants). [papadimitriou-etal-2025-vlm-linear-structure]

models: AIMv2 · method: BatchTopK Sparse Autoencoders, Geometric analysis

Mamba

Lost in State Space: Probing Frozen Mamba Representations (2026)measured

Frozen Mamba collapses into an extreme anisotropy cone

Details

- Frozen Mamba-130m activations collapse into an extremely narrow anisotropy cone: mean pairwise cosine similarity 0.9999 (std 4.4e-5), against RoBERTa-base's ~0.9839. [wagh-singh-2026-lost-in-state-space-probing-frozen-mamba] - The collapse cripples frozen probes (CoLA MCC 0.000 for the raw final state; patched-Mamba 0.246 vs RoBERTa 0.475; STS-B 0.316 vs 0.650). [wagh-singh-2026-lost-in-state-space-probing-frozen-mamba] - A geometry-targeted orthogonal write-injection (eta=0.5) fails to widen the cone and worsens STS-B (0.316 to 0.259), a candid negative result implying geometry correction needs training-time adaptation. [wagh-singh-2026-lost-in-state-space-probing-frozen-mamba] - The comparison model is RoBERTa-base, not BERT, and the "~10x-tighter" figure is the angular ratio between cosine 0.9999 and 0.99, not a measured Mamba-vs-baseline cone ratio. [wagh-singh-2026-lost-in-state-space-probing-frozen-mamba]

models: Mamba-130m (state-spaces/mamba-130m-hf) · method: Linear probing, Causal interventions (steering)

GPT-OSS

The Myth of Expert Specialization in MoEs: Why Routing Reflects Geometry, Not Necessarily Domain Expertise (2026)measured

MoE routing reflects hidden-state geometry, not domain expertise

Details

- Under the load-balancing auxiliary loss, MoE routers are pushed to suppress the shared dominant direction of their hidden states (Proposition 2), so routing tracks hidden-state geometry rather than domain content. [wang-etal-2026-myth-of-expert-specialization] - Hidden-state cosine similarity predicts expert-usage similarity, and two different models solving the same math problem share ~60% of their most-used experts, the same overlap as one model across different questions. [wang-etal-2026-myth-of-expert-specialization] - Prefill expert usage is nearly identical across semantically different inputs and diverges only during generation, with deep-layer router collapse. [wang-etal-2026-myth-of-expert-specialization] - Studied on five production MoE models: gpt-oss-20B, ERNIE-4.5-21B-A3B-Base, Qwen-3-30B, Ling-mini and Trinity-Mini-Base (Moonlight-16B is a comparison only); observational. [wang-etal-2026-myth-of-expert-specialization]

models: GPT-OSS-20B · method: SVD, Geometric analysis

ERNIE

The Myth of Expert Specialization in MoEs: Why Routing Reflects Geometry, Not Necessarily Domain Expertise (2026)measured

MoE routing reflects hidden-state geometry, not domain expertise

Details

- Under the load-balancing auxiliary loss, MoE routers are pushed to suppress the shared dominant direction of their hidden states (Proposition 2), so routing tracks hidden-state geometry rather than domain content. [wang-etal-2026-myth-of-expert-specialization] - Hidden-state cosine similarity predicts expert-usage similarity, and two different models solving the same math problem share ~60% of their most-used experts, the same overlap as one model across different questions. [wang-etal-2026-myth-of-expert-specialization] - Prefill expert usage is nearly identical across semantically different inputs and diverges only during generation, with deep-layer router collapse. [wang-etal-2026-myth-of-expert-specialization] - Studied on five production MoE models: gpt-oss-20B, ERNIE-4.5-21B-A3B-Base, Qwen-3-30B, Ling-mini and Trinity-Mini-Base (Moonlight-16B is a comparison only); observational. [wang-etal-2026-myth-of-expert-specialization]

models: ERNIE-4.5-21B-A3B-Base · method: SVD, Geometric analysis

Ling

The Myth of Expert Specialization in MoEs: Why Routing Reflects Geometry, Not Necessarily Domain Expertise (2026)measured

MoE routing reflects hidden-state geometry, not domain expertise

Details

- Under the load-balancing auxiliary loss, MoE routers are pushed to suppress the shared dominant direction of their hidden states (Proposition 2), so routing tracks hidden-state geometry rather than domain content. [wang-etal-2026-myth-of-expert-specialization] - Hidden-state cosine similarity predicts expert-usage similarity, and two different models solving the same math problem share ~60% of their most-used experts, the same overlap as one model across different questions. [wang-etal-2026-myth-of-expert-specialization] - Prefill expert usage is nearly identical across semantically different inputs and diverges only during generation, with deep-layer router collapse. [wang-etal-2026-myth-of-expert-specialization] - Studied on five production MoE models: gpt-oss-20B, ERNIE-4.5-21B-A3B-Base, Qwen-3-30B, Ling-mini and Trinity-Mini-Base (Moonlight-16B is a comparison only); observational. [wang-etal-2026-myth-of-expert-specialization]

models: Ling-mini-Base · method: SVD, Geometric analysis

Trinity

The Myth of Expert Specialization in MoEs: Why Routing Reflects Geometry, Not Necessarily Domain Expertise (2026)measured

MoE routing reflects hidden-state geometry, not domain expertise

Details

- Under the load-balancing auxiliary loss, MoE routers are pushed to suppress the shared dominant direction of their hidden states (Proposition 2), so routing tracks hidden-state geometry rather than domain content. [wang-etal-2026-myth-of-expert-specialization] - Hidden-state cosine similarity predicts expert-usage similarity, and two different models solving the same math problem share ~60% of their most-used experts, the same overlap as one model across different questions. [wang-etal-2026-myth-of-expert-specialization] - Prefill expert usage is nearly identical across semantically different inputs and diverges only during generation, with deep-layer router collapse. [wang-etal-2026-myth-of-expert-specialization] - Studied on five production MoE models: gpt-oss-20B, ERNIE-4.5-21B-A3B-Base, Qwen-3-30B, Ling-mini and Trinity-Mini-Base (Moonlight-16B is a comparison only); observational. [wang-etal-2026-myth-of-expert-specialization]

models: Trinity-Mini-Base · method: SVD, Geometric analysis

LaBSE

WhiteningBERT: An Easy Unsupervised Sentence Embedding Approach (2021)measured

Whitening removes anisotropy and improves sentence similarity

Details

- Whitening, a parameter-free linear map (mean-center, then rotate and scale by the covariance eigendecomposition), makes sentence embeddings isotropic and improves cosine STS. [huang-etal-2021] - Average STS Spearman rises 62.97 to 67.76 (BERT), 59.53 to 67.72 (RoBERTa) and 64.12 to 68.30 (DistilBERT), but only 71.56 to 71.71 for the already-isotropic LaBSE. [huang-etal-2021] - Averaging the first and last layers (L1+L12) beats any single layer, and token-averaging beats the [CLS] vector by a large margin. [huang-etal-2021] - Main models are BERT-base, RoBERTa-base, DistilBERT and LaBSE, with the effect confirmed across 24 pretrained encoders in the appendix. [huang-etal-2021]

models: LaBSE · method: Whitening transformation

NanoGPT Causal-Attention, No Positional Encoding (Causal-NoPE)

Position Information Emerges in Causal Transformers Without Positional Encodings via Similarity of Nearby Embeddings (2025)measured

Causal attention without positions creates adjacency decay

Details

- In causal-attention transformers with no positional encodings, the self-cosine-similarity matrix shows a distance-decaying adjacency pattern: nearby positions are more similar than distant ones. [zuo-etal-2025-position-information-emerges-causal-transformers-similarity-nearby-embeddings] - An adjacency-probability score jumps from ~0.48-0.58 at the raw token-embedding layer to 0.80-0.99 after the first causal-attention layer, and appears even in randomly initialized models. [zuo-etal-2025-position-information-emerges-causal-transformers-similarity-nearby-embeddings] - The pattern is absent under non-causal attention (~0.5) and follows from causal prefix-averaging plus the near-orthogonality of high-dimensional embeddings. [zuo-etal-2025-position-information-emerges-causal-transformers-similarity-nearby-embeddings] - Trained NanoGPT models (6/12/24 layers, hidden 192/384/768) on four synthetic position-sensitive tasks; no pretrained LLMs. [zuo-etal-2025-position-information-emerges-causal-transformers-similarity-nearby-embeddings]

models: NanoGPT Causal-NoPE, 6-layer (10.6M params) · method: Geometric analysis