MATH · IN · MODELS
structures / Manifolds / Representational-similarity trajectory across depth

Representational-similarity trajectory across depth

PROPERTYmeasurementfunctionalintermediatehow it's classified →

How similar a network's own layer-wise representations are to each other (via a similarity metric such as CCA, SVCCA, or CKA) traces a specific, measurable trajectory across depth — layers cluster into similarity-based stages, and the trajectory's shape shifts with training objective, architecture, or between independently trained networks.

Replicationcomputed from the corpus — never hand-assigned
8 papers · no shared authors7 architecture classes · across papers3 domains · across papers16 model families · across papers
Filled = two or more values reported by papers that share no author — replication. Outlined = two or more values, but all from a single study — breadth, not replication. Grey = a single value. Derived from paper authorship and each model's architecture class, domain and family; it updates itself when a paper is added.

Definition

Given a similarity functional ρ(X,Y)[0,1]\rho(X,Y)\in[0,1] between two representation matrices (rows = examples, columns = neurons/directions) that is invariant to the specific basis each layer happens to use (e.g. canonical correlation analysis (CCA), SVCCA’s SVD-then-CCA composition, or centered kernel alignment (CKA)), this concept describes the trajectory ρ(X,X)\ell \mapsto \rho(X_\ell, X_{\ell'}) traced out as \ell ranges over a network’s own layers — either compared to a fixed reference layer (e.g. the final layer, or a converged/late-training checkpoint), or pairwise across all layers, or across two independently trained networks’ corresponding layers. Unlike Intrinsic-dimension profile across depth (which tracks a single scalar, the local dimensionality, layer by layer), this concept tracks a relational quantity — how alike two representations are — and the resulting curve/matrix is itself the measured geometric object.

Intuition

Two networks (or two layers of the same network) can each individually look like a normal, reasonably-shaped point cloud, and yet be almost unrelated to each other, or nearly identical up to a change of basis. A similarity trajectory answers “which layers/networks are doing functionally the same thing, geometrically speaking” without requiring either representation to already be aligned to a common coordinate system — which is exactly why CCA-family metrics (invariant to invertible linear transformation) and CKA (invariant to orthogonal transformation and isotropic scaling) are preferred over raw distance comparisons here.

Properties

  • A trajectory, not a single number. The object of interest is the whole curve (or similarity matrix) across depth, not one similarity score — e.g. a curve that stays flat then drops sharply at one layer indicates a genuine architectural transition, which a single averaged similarity score would wash out.

  • Basis-invariant by construction. Because CCA/SVCCA/CKA are invariant to (at least) invertible linear reparameterization of each representation, a low similarity score cannot be explained away as “the two layers just use different coordinate systems for the same information” — genuine representational divergence is what remains once this invariance is accounted for.

  • Distinct from intrinsic-dimension-profile. The dimension profile asks “how many effective coordinates does this one representation need,” entirely intrinsic to a single layer; this concept asks “how related are two representations to each other,” which requires no ID estimation at all and can reveal structure (e.g. two very-differently-shaped-but-highly-correlated layers) invisible to a dimension-only analysis.

  • Foundational precedent: an information-theoretic bottom-up evolution of feature geometry, tracked across depth and training objective. Voita, Sennrich & Titov (2019) trace CCA similarity (against each Transformer’s own final layer) and mutual-information estimates across depth in real trained machine-translation, left-to-right language-model, and masked-language-model Transformers, finding a training-objective-dependent trajectory: LM representations progressively discard “past” token information while building toward the “future” prediction target, while MLM representations first generalize away from surface token identity (peak abstraction in middle layers) before partially reconstructing token identity again near the output. See voita-etal-2019-cca-and-mutual-information-trajectories-reveal-a-training-objective-dependent-bottom-up-evolution-of-transformer-representation-geometry-across-depth.

  • The originating instance: an affine-invariant similarity metric (SVD + CCA) reveals substantial over-parameterization and a bottom-up training-convergence pattern in real trained CNNs and RNNs. Raghu, Gilmer, Yosinski & Sohl-Dickstein (2017, SVCCA) measure effective dimensionality per layer (several layers usable at far lower effective dimension than their nominal width) and show earlier layers stabilize earlier during training than later layers — the founding instance of the CCA-based similarity-trajectory family later extended by Voita et al. (2019, above). See raghu-etal-2017-svcca-affine-invariant-similarity-reveals-representation-convergence-dynamics-and-effective-dimensionality-in-real-trained-networks.

  • A distinct similarity index — invariant to orthogonal transformation and scaling rather than to any invertible linear transform — reliably recovers known correspondences where the CCA family fails. Kornblith, Norouzi, Lee & Hinton (2019, CKA) show CCA-family metrics become unreliable when representation dimension is large relative to sample count, and that Centered Kernel Alignment instead reliably recovers known layer/network correspondences across real trained CNNs of varying width, depth, and initialization. See kornblith-etal-2019-cka-reliably-recovers-known-layer-and-network-correspondences-in-real-trained-cnns-where-cca-based-metrics-fail.

  • A projection-weighted refinement of CCA fixes its sensitivity to low-variance directions, revealing that wider real networks converge to more similar representations and generalizing networks are more mutually similar than memorizing ones. Morcos, Raghu & Bengio (2018, PWCCA) weight each canonical-correlation direction by how much of the original representation it explains, and apply this to real CIFAR-10 convnets and PTB/WikiText-2 LSTM language models across many independent training runs — similarity between independently-trained-but-identical architectures increases with network width, and networks trained on real vs. randomly-shuffled labels (generalizing vs. memorizing solutions) are measurably more self-similar within each group than across groups. See morcos-etal-2018-pwcca-shows-wider-networks-converge-to-more-similar-real-representations-and-generalizing-networks-are-more-similar-to-each-other-than-to-memorizing-ones.

  • VIRAL: MLLM internal visual-token representations diverge from frozen vision-foundation-model features across depth, and an explicit alignment loss causally recovers benchmark performance. Yoon, Jung, Kim et al. (2025) measure CKNNA similarity between real LLaVA-1.5-7B/13B’s internal visual-token representations and frozen VFM (DINOv2/CLIP/SAM/etc.) features across all 32 layers, finding a sharp early-layer drop with only partial mid-layer recovery; adding a training-time cosine-similarity alignment loss causally improves MMVP (28.20%->33.33%), CV-Bench2D (56.82%->59.67%), and POPE (85.70%->88.32%) for LLaVA-1.5-7B+DINOv2. See yoon-etal-2025-viral-mllm-internal-visual-token-representations-diverge-from-frozen-vision-foundation-model-features-across-depth-and-an-explicit-alignment-loss-causally-recovers-benchmark-performance.

  • Trajectory length and curvature, as distinct from a relational similarity curve. Pandey, Singh & Mahdid (2026) measure a related but distinct property of the same layer-by-layer evolution: not how similar two layers’ representations are to each other, but the geometric shape (length, curvature) of the path a population of representations itself traces through ambient space as it moves from layer to layer, across three real trained transformers (GPT-2, TinyLlama, Qwen2.5). Reasoning/analogy prompts produce trajectories of significantly greater curvature than lexical-variation prompts (0.71-0.83 rad vs. 0.27-0.31 rad); semantically related prompts show trajectory convergence peaking at middle-to-late layers (convergence index 0.41-0.58, p<0.001); ambiguous tokens show measurable trajectory bifurcation (up to 5.6x final-layer separation). A layerwise-cosine-similarity component of the same analysis recovers a universal three-phase (encoding/elaboration/output-preparation) structure consistent across all three architectures — this component is the closest to a conventional similarity-trajectory measurement, while trajectory length/curvature is a genuinely distinct, complementary geometric quantity about the path’s own shape rather than a pairwise relational score. Purely observational. See pandey-etal-2026-real-gpt-2-tinyllama-and-qwen2-5-representation-trajectories-show-higher-curvature-for-reasoning-than-lexical-tasks-and-a-universal-three-phase-layerwise-structure.

  • Token-level trajectory curvature and layer-wise entropy across real speech-SSL models, distinguishing distinct optimization regimes. Sadok & Alameda-Pineda (2026, InsideSSL) apply the same trajectory-curvature instrument as Pandey et al. above (curvature of the path traced by adjacent-token transition vectors), but across layers of real pretrained Wav2Vec2, WavLM, HuBERT, and Data2Vec-audio speech models on real LibriSpeech audio: curvature starts high (about 1.4) and decreases to stabilize around 1.2 with depth for most models (manifold “unfolding” toward linear separability), while paired per-layer entropy measurements reveal Wav2Vec2 alone undergoes a sharp entropy collapse at layer 11, coinciding with a spike in an InfoNCE-based invariance-loss measurement — a training-objective-dependent divergence in trajectory shape, paralleling the MT/LM/MLM divergence Voita et al. found in text transformers above, but for the compression (entropy) dimension of the trajectory rather than curvature alone (see also Dimensional collapse). See sadok-alameda-pineda-2026-real-speech-ssl-models-wav2vec2-wavlm-hubert-and-data2vec-show-distinct-layerwise-entropy-and-trajectory-curvature-regimes-with-wav2vec2-undergoing-a-late-layer-entropy-collapse.

  • A “tubularity” trajectory-shape metric (tightness and self-crossings of the layer-to-layer path), applied to a real artificial neural network trained to predict biological visual-cortex responses, distinguishes module-specific trajectory geometry and finds artificial trajectories reorganize more (fewer crossings) than the biological data used only as an external comparison baseline. Bertram, Dyballa, Keller, Kinger & Zucker (2026) build both decoding manifolds (PCA of stimulus-averaged responses) and encoding manifolds (nonnegative tensor factorization plus diffusion maps of neuron x stimulus x time responses) from a real trained foundation model of neural activity (Wang et al.’s FNN, a DenseNet-conv-encoder plus convLSTM plus linear-readout network trained on real mouse visual-cortex MICrONS recordings), finding decoding accuracy rises from 0.59 (early conv encoder) to 0.89 (recurrent module) to 0.88 (readout), while the tubularity metric shows the recurrent module actively “pushes apart” stimulus trajectories (tightness 0.1228, crossings 2.65e-7) with significantly fewer crossings than the biological retina/V1 comparison data (p<0.005, Bonferroni) — a genuinely different trajectory-shape instrument (tightness/self-crossing rather than curvature or CCA/CKA similarity), applied to a scientific/biological-prediction ANN rather than an LLM. See bertram-etal-2026-a-tubularity-trajectory-shape-metric-in-a-real-ann-trained-to-predict-mouse-visual-cortex-responses-finds-the-recurrent-module-actively-separates-stimulus-trajectories-with-fewer-crossings-than-biological-data.

Exercises

Base

  1. Why is CCA (or CKA), rather than raw Euclidean distance between two weight-aligned representations, the right tool for comparing two different layers’ representations?
Solution

Two layers can encode identical information while using entirely different, arbitrarily-rotated-and-rescaled coordinate systems for their neurons (there is no reason different layers, let alone different networks, should share a common “meaning” per coordinate axis). Raw Euclidean distance is sensitive to exactly this kind of basis change — it would report two functionally identical representations as maximally dissimilar if their bases differ. CCA and CKA are constructed specifically to be invariant to such reparameterizations (CCA to any invertible linear change of basis, CKA to orthogonal transformations and isotropic scaling), so the reported similarity reflects the actual shared information content rather than an artifact of coordinate choice.

Middle

  1. A similarity trajectory (against the final layer) is flat and high for a network’s first three layers, then drops sharply at layer four before climbing back to a high value by the final layer. What would this trajectory shape suggest about layer four’s role, and how would you distinguish this from the same phenomenon appearing simply because layer four’s representation has an unusually different intrinsic dimension?
Solution

A sharp mid-trajectory dip suggests layer four performs some transformation that substantially reorganizes the representation relative to both its neighbors and the eventual output — a genuine “bottleneck” or transitional stage, consistent with the MLM finding above (early generalization away from surface form). To rule out that this is simply an artifact of layer four having a very different intrinsic dimension (which could depress a similarity score for reasons unrelated to genuine representational change), one would consult the layer’s own Intrinsic-dimension profile across depth value independently — if the ID profile shows nothing unusual at layer four, the similarity dip is better attributed to a genuine content/organization shift rather than a dimension mismatch.

Found in (8 observations · 15 families)

FNN (Foundation Model of Neural Activity)

Manifolds and Modules: How Function Develops in a Neural Foundation Model (2026)measured

A tubularity trajectory-shape metric distinguishes module-specific geometry in a real trained neural-activity foundation model

Details

Bertram, Dyballa, Keller, Kinger & Zucker build decoding manifolds (PCA of stimulus-averaged responses) and encoding manifolds (tensor factorization plus diffusion maps) from Wang et al.'s real trained FNN, a DenseNet-conv-encoder plus convLSTM plus linear-readout network fit to predict real mouse visual-cortex (MICrONS) responses [bertram-etal-2026-manifolds-and-modules-neural-foundation-model] Layer-wise decoding accuracy rises from 0.59 (early convolutional encoder) to 0.89 (recurrent module) to 0.88 (readout) [bertram-etal-2026-manifolds-and-modules-neural-foundation-model] A novel tubularity metric (trajectory tightness and self-crossings across layers) shows the recurrent module actively pushes apart stimulus trajectories, with significantly fewer crossings than the biological retina/V1 comparison data (p<0.005, Bonferroni-corrected) [bertram-etal-2026-manifolds-and-modules-neural-foundation-model]

models: FNN (DenseNet-conv encoder + convLSTM + linear readout, MICrONS mouse V1) · method:

Convolutional neural network (generic)

Similarity of Neural Network Representations Revisited (2019)measured

CKA recovers layer and network correspondences where CCA metrics fail

Details

Kornblith et al. show CCA-family similarity metrics, invariant to any invertible linear reparameterization, become unreliable when representation dimension is large relative to sample count [kornblith-etal-2019-cka-similarity-revisited] They introduce Centered Kernel Alignment (CKA), invariant only to orthogonal transformation and isotropic scaling [kornblith-etal-2019-cka-similarity-revisited] On sanity checks with Tiny-10 and Plain-(8n+2) CNNs, plus ResNet and Transformer architectures, CKA reliably recovers known layer/network correspondences (e.g. matching layers across different random seeds) that SVCCA and other CCA-based metrics fail to recover [kornblith-etal-2019-cka-similarity-revisited] The comparison is observational, with no causal intervention [kornblith-etal-2019-cka-similarity-revisited]

models: Convolutional network image classifier (generic) · method: Centered Kernel Alignment (CKA) representation similarity
SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability (2017)measured

SVCCA reveals low effective dimensionality and bottom-up convergence

Details

Raghu et al. introduce SVCCA (SVD truncation followed by CCA) and apply it to trained CNNs (a ResNet and a plain convnet on CIFAR-10/ImageNet) and RNNs [raghu-etal-2017-svcca] Individual layers are usable at far lower effective dimension than their nominal unit count, evidencing substantial over-parameterization [raghu-etal-2017-svcca] Representations converge bottom-up: earlier layers stabilize earlier during training than later layers [raghu-etal-2017-svcca] This is the foundational CCA-based similarity method later extended by Voita et al. and contrasted against by CKA [raghu-etal-2017-svcca]

models: Convolutional network image classifier (generic) · method: SVCCA (SVD + canonical correlation) representation similarity

ResNet

Similarity of Neural Network Representations Revisited (2019)measured

CKA recovers layer and network correspondences where CCA metrics fail

Details

Kornblith et al. show CCA-family similarity metrics, invariant to any invertible linear reparameterization, become unreliable when representation dimension is large relative to sample count [kornblith-etal-2019-cka-similarity-revisited] They introduce Centered Kernel Alignment (CKA), invariant only to orthogonal transformation and isotropic scaling [kornblith-etal-2019-cka-similarity-revisited] On sanity checks with Tiny-10 and Plain-(8n+2) CNNs, plus ResNet and Transformer architectures, CKA reliably recovers known layer/network correspondences (e.g. matching layers across different random seeds) that SVCCA and other CCA-based metrics fail to recover [kornblith-etal-2019-cka-similarity-revisited] The comparison is observational, with no causal intervention [kornblith-etal-2019-cka-similarity-revisited]

models: ResNet (image classifier, various depths) · method: Centered Kernel Alignment (CKA) representation similarity
SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability (2017)measured

SVCCA reveals low effective dimensionality and bottom-up convergence

Details

Raghu et al. introduce SVCCA (SVD truncation followed by CCA) and apply it to trained CNNs (a ResNet and a plain convnet on CIFAR-10/ImageNet) and RNNs [raghu-etal-2017-svcca] Individual layers are usable at far lower effective dimension than their nominal unit count, evidencing substantial over-parameterization [raghu-etal-2017-svcca] Representations converge bottom-up: earlier layers stabilize earlier during training than later layers [raghu-etal-2017-svcca] This is the foundational CCA-based similarity method later extended by Voita et al. and contrasted against by CKA [raghu-etal-2017-svcca]

models: ResNet (image classifier, various depths) · method: SVCCA (SVD + canonical correlation) representation similarity

Transformer (generic)

Similarity of Neural Network Representations Revisited (2019)measured

CKA recovers layer and network correspondences where CCA metrics fail

Details

Kornblith et al. show CCA-family similarity metrics, invariant to any invertible linear reparameterization, become unreliable when representation dimension is large relative to sample count [kornblith-etal-2019-cka-similarity-revisited] They introduce Centered Kernel Alignment (CKA), invariant only to orthogonal transformation and isotropic scaling [kornblith-etal-2019-cka-similarity-revisited] On sanity checks with Tiny-10 and Plain-(8n+2) CNNs, plus ResNet and Transformer architectures, CKA reliably recovers known layer/network correspondences (e.g. matching layers across different random seeds) that SVCCA and other CCA-based metrics fail to recover [kornblith-etal-2019-cka-similarity-revisited] The comparison is observational, with no causal intervention [kornblith-etal-2019-cka-similarity-revisited]

models: Transformer image classifier (generic) · method: Centered Kernel Alignment (CKA) representation similarity

Custom Research CNN (purpose-built convolutional net for representation-geometry studies)

Insights on Representational Similarity in Neural Networks with Canonical Correlation Analysis (2018)measured

Wider networks converge to more similar representations, measured by PWCCA

Details

Morcos et al. introduce projection-weighted CCA (PWCCA), weighting each canonical correlation direction by how much of the original representation it explains, fixing SVCCA's sensitivity to low-variance directions [morcos-etal-2018-pwcca] Applied to CIFAR-10 convnets and PTB/WikiText-2 LSTMs across many runs, wider networks qualitatively converge to more similar solutions across random seeds (no coefficient is attached to the width relationship) [morcos-etal-2018-pwcca] Test accuracy is strongly anti-correlated with pairwise CCA distance (-0.96), so networks closer to the convergent solution generalize better [morcos-etal-2018-pwcca] Generalizing networks converge to more similar solutions, but memorizing networks do NOT self-cluster: they are as similar to each other as they are to a generalizing network [morcos-etal-2018-pwcca]

models: 11-layer CIFAR-10 convnet (Morcos et al. 2018) · method: Projection-Weighted CCA (PWCCA) representation similarity

RNNLM (Recurrent Neural Network Language Model)

Insights on Representational Similarity in Neural Networks with Canonical Correlation Analysis (2018)measured

Wider networks converge to more similar representations, measured by PWCCA

Details

Morcos et al. introduce projection-weighted CCA (PWCCA), weighting each canonical correlation direction by how much of the original representation it explains, fixing SVCCA's sensitivity to low-variance directions [morcos-etal-2018-pwcca] Applied to CIFAR-10 convnets and PTB/WikiText-2 LSTMs across many runs, wider networks qualitatively converge to more similar solutions across random seeds (no coefficient is attached to the width relationship) [morcos-etal-2018-pwcca] Test accuracy is strongly anti-correlated with pairwise CCA distance (-0.96), so networks closer to the convergent solution generalize better [morcos-etal-2018-pwcca] Generalizing networks converge to more similar solutions, but memorizing networks do NOT self-cluster: they are as similar to each other as they are to a generalizing network [morcos-etal-2018-pwcca]

models: LSTM language model (AWD-LSTM setup, PTB/WikiText-2) · method: Projection-Weighted CCA (PWCCA) representation similarity

GPT-2

Trajectory Geometry of Transformer Representations Across Layers (2026)measured

Real GPT-2, TinyLlama, and Qwen2.5 representation trajectories show higher curvature for reasoning than lexical tasks, and a universal three-phase layerwise structure

Details

Five geometric metrics (trajectory length, curvature, semantic convergence index, layerwise cosine similarity, representational stability) are computed on the population trajectory traced by hidden representations across layers of three real trained transformers -- GPT-2, TinyLlama, and Qwen2.5 -- over five semantically controlled prompt families [pandey-etal-2026-trajectory-geometry-transformer-representations] Reasoning and analogy tasks produce trajectories of significantly greater curvature than lexical-variation tasks (0.71-0.83 rad vs. 0.27-0.31 rad) across all three architectures [pandey-etal-2026-trajectory-geometry-transformer-representations] Semantically related prompts show statistically significant trajectory convergence peaking in middle-to-late layers (convergence index 0.41-0.58, p<0.001, Mann-Whitney U), and ambiguous tokens show measurable trajectory bifurcation of up to 5.6x final-layer separation, absent in unambiguous controls [pandey-etal-2026-trajectory-geometry-transformer-representations] Layerwise cosine similarity reveals a universal three-phase structure (encoding, elaboration, output preparation) whose boundaries are consistent across all three architectures [pandey-etal-2026-trajectory-geometry-transformer-representations]

models: GPT-2 Small · method:

TinyLlama

Trajectory Geometry of Transformer Representations Across Layers (2026)measured

Real GPT-2, TinyLlama, and Qwen2.5 representation trajectories show higher curvature for reasoning than lexical tasks, and a universal three-phase layerwise structure

Details

Five geometric metrics (trajectory length, curvature, semantic convergence index, layerwise cosine similarity, representational stability) are computed on the population trajectory traced by hidden representations across layers of three real trained transformers -- GPT-2, TinyLlama, and Qwen2.5 -- over five semantically controlled prompt families [pandey-etal-2026-trajectory-geometry-transformer-representations] Reasoning and analogy tasks produce trajectories of significantly greater curvature than lexical-variation tasks (0.71-0.83 rad vs. 0.27-0.31 rad) across all three architectures [pandey-etal-2026-trajectory-geometry-transformer-representations] Semantically related prompts show statistically significant trajectory convergence peaking in middle-to-late layers (convergence index 0.41-0.58, p<0.001, Mann-Whitney U), and ambiguous tokens show measurable trajectory bifurcation of up to 5.6x final-layer separation, absent in unambiguous controls [pandey-etal-2026-trajectory-geometry-transformer-representations] Layerwise cosine similarity reveals a universal three-phase structure (encoding, elaboration, output preparation) whose boundaries are consistent across all three architectures [pandey-etal-2026-trajectory-geometry-transformer-representations]

models: TinyLlama-1.1B · method:

Qwen

Trajectory Geometry of Transformer Representations Across Layers (2026)measured

Real GPT-2, TinyLlama, and Qwen2.5 representation trajectories show higher curvature for reasoning than lexical tasks, and a universal three-phase layerwise structure

Details

Five geometric metrics (trajectory length, curvature, semantic convergence index, layerwise cosine similarity, representational stability) are computed on the population trajectory traced by hidden representations across layers of three real trained transformers -- GPT-2, TinyLlama, and Qwen2.5 -- over five semantically controlled prompt families [pandey-etal-2026-trajectory-geometry-transformer-representations] Reasoning and analogy tasks produce trajectories of significantly greater curvature than lexical-variation tasks (0.71-0.83 rad vs. 0.27-0.31 rad) across all three architectures [pandey-etal-2026-trajectory-geometry-transformer-representations] Semantically related prompts show statistically significant trajectory convergence peaking in middle-to-late layers (convergence index 0.41-0.58, p<0.001, Mann-Whitney U), and ambiguous tokens show measurable trajectory bifurcation of up to 5.6x final-layer separation, absent in unambiguous controls [pandey-etal-2026-trajectory-geometry-transformer-representations] Layerwise cosine similarity reveals a universal three-phase structure (encoding, elaboration, output preparation) whose boundaries are consistent across all three architectures [pandey-etal-2026-trajectory-geometry-transformer-representations]

models: Qwen2.5-7B · method:

wav2vec 2.0

InsideSSL: Understanding Self-Supervised Speech Representations using a Model-Centric Perspective (2026)measured

Real speech-SSL models Wav2Vec2, WavLM, HuBERT, and Data2Vec show distinct layer-wise entropy and trajectory-curvature regimes, with Wav2Vec2 undergoing a late-layer entropy collapse

Details

Von Neumann entropy (from the per-layer Gram matrix) and average curvature of the trajectory traced by adjacent-token transition vectors are computed layer-by-layer for real pretrained Wav2Vec2-base, WavLM-base, HuBERT-base, and Data2Vec-audio-base on the real LibriSpeech test-clean corpus [sadok-alameda-pineda-2026-insidessl] WavLM, HuBERT, and Data2Vec sustain high normalized entropy throughout the network (about 0.82 declining only to about 0.75 in final layers), while Wav2Vec2-base instead undergoes an entropy collapse toward near-baseline levels at layer 11, coinciding with a sharp rise in InfoNCE-based invariance loss (average 1.0 to 3.0) across its final two layers [sadok-alameda-pineda-2026-insidessl] Trajectory curvature starts high (about 1.4) in early layers and decreases to stabilize around 1.2 in deeper layers for most models, indicating the embedding manifold "unfolds" toward linear separability with depth; WavLM, HuBERT and UniSpeech show highly correlated curvature trajectories (Pearson correlation above 0.96), while linear probing shows phoneme-classification accuracy peaks at the depth where curvature transitions from the high- to low-curvature regime [sadok-alameda-pineda-2026-insidessl]

models: wav2vec2-base-960h · method:

WavLM

InsideSSL: Understanding Self-Supervised Speech Representations using a Model-Centric Perspective (2026)measured

Real speech-SSL models Wav2Vec2, WavLM, HuBERT, and Data2Vec show distinct layer-wise entropy and trajectory-curvature regimes, with Wav2Vec2 undergoing a late-layer entropy collapse

Details

Von Neumann entropy (from the per-layer Gram matrix) and average curvature of the trajectory traced by adjacent-token transition vectors are computed layer-by-layer for real pretrained Wav2Vec2-base, WavLM-base, HuBERT-base, and Data2Vec-audio-base on the real LibriSpeech test-clean corpus [sadok-alameda-pineda-2026-insidessl] WavLM, HuBERT, and Data2Vec sustain high normalized entropy throughout the network (about 0.82 declining only to about 0.75 in final layers), while Wav2Vec2-base instead undergoes an entropy collapse toward near-baseline levels at layer 11, coinciding with a sharp rise in InfoNCE-based invariance loss (average 1.0 to 3.0) across its final two layers [sadok-alameda-pineda-2026-insidessl] Trajectory curvature starts high (about 1.4) in early layers and decreases to stabilize around 1.2 in deeper layers for most models, indicating the embedding manifold "unfolds" toward linear separability with depth; WavLM, HuBERT and UniSpeech show highly correlated curvature trajectories (Pearson correlation above 0.96), while linear probing shows phoneme-classification accuracy peaks at the depth where curvature transitions from the high- to low-curvature regime [sadok-alameda-pineda-2026-insidessl]

models: WavLM Base · method:

HuBERT

InsideSSL: Understanding Self-Supervised Speech Representations using a Model-Centric Perspective (2026)measured

Real speech-SSL models Wav2Vec2, WavLM, HuBERT, and Data2Vec show distinct layer-wise entropy and trajectory-curvature regimes, with Wav2Vec2 undergoing a late-layer entropy collapse

Details

Von Neumann entropy (from the per-layer Gram matrix) and average curvature of the trajectory traced by adjacent-token transition vectors are computed layer-by-layer for real pretrained Wav2Vec2-base, WavLM-base, HuBERT-base, and Data2Vec-audio-base on the real LibriSpeech test-clean corpus [sadok-alameda-pineda-2026-insidessl] WavLM, HuBERT, and Data2Vec sustain high normalized entropy throughout the network (about 0.82 declining only to about 0.75 in final layers), while Wav2Vec2-base instead undergoes an entropy collapse toward near-baseline levels at layer 11, coinciding with a sharp rise in InfoNCE-based invariance loss (average 1.0 to 3.0) across its final two layers [sadok-alameda-pineda-2026-insidessl] Trajectory curvature starts high (about 1.4) in early layers and decreases to stabilize around 1.2 in deeper layers for most models, indicating the embedding manifold "unfolds" toward linear separability with depth; WavLM, HuBERT and UniSpeech show highly correlated curvature trajectories (Pearson correlation above 0.96), while linear probing shows phoneme-classification accuracy peaks at the depth where curvature transitions from the high- to low-curvature regime [sadok-alameda-pineda-2026-insidessl]

models: HuBERT-base · method:

Data2Vec

InsideSSL: Understanding Self-Supervised Speech Representations using a Model-Centric Perspective (2026)measured

Real speech-SSL models Wav2Vec2, WavLM, HuBERT, and Data2Vec show distinct layer-wise entropy and trajectory-curvature regimes, with Wav2Vec2 undergoing a late-layer entropy collapse

Details

Von Neumann entropy (from the per-layer Gram matrix) and average curvature of the trajectory traced by adjacent-token transition vectors are computed layer-by-layer for real pretrained Wav2Vec2-base, WavLM-base, HuBERT-base, and Data2Vec-audio-base on the real LibriSpeech test-clean corpus [sadok-alameda-pineda-2026-insidessl] WavLM, HuBERT, and Data2Vec sustain high normalized entropy throughout the network (about 0.82 declining only to about 0.75 in final layers), while Wav2Vec2-base instead undergoes an entropy collapse toward near-baseline levels at layer 11, coinciding with a sharp rise in InfoNCE-based invariance loss (average 1.0 to 3.0) across its final two layers [sadok-alameda-pineda-2026-insidessl] Trajectory curvature starts high (about 1.4) in early layers and decreases to stabilize around 1.2 in deeper layers for most models, indicating the embedding manifold "unfolds" toward linear separability with depth; WavLM, HuBERT and UniSpeech show highly correlated curvature trajectories (Pearson correlation above 0.96), while linear probing shows phoneme-classification accuracy peaks at the depth where curvature transitions from the high- to low-curvature regime [sadok-alameda-pineda-2026-insidessl]

models: Data2Vec Audio Base (LibriSpeech-960h) · method:

Voita et al. (2019) objective-matched Transformers (MT/LM/MLM)

The Bottom-up Evolution of Representations in the Transformer: A Study with Machine Translation and Language Modeling Objectives (2019)measured

CCA and mutual-information trajectories differ by training objective across depth

Details

Voita et al. track layer-wise CCA similarity (against each model's own final layer) and mutual information with past/future tokens and surface identity across three training objectives (MT, LM, MLM) [voita-etal-2019-bottom-up-evolution-of-representations] LM representations show monotonically increasing divergence from early layers, discarding past-token information while building toward the future-prediction target [voita-etal-2019-bottom-up-evolution-of-representations] MLM representations first move away from surface token identity in early-to-middle layers, then partially reconstruct it near the output [voita-etal-2019-bottom-up-evolution-of-representations] The trajectory shape is thus systematically training-objective-dependent; no causal intervention is performed [voita-etal-2019-bottom-up-evolution-of-representations]

models: Transformer-base trained on machine translation (En-Ru/En-Fr), Transformer-base trained as a left-to-right language model, Transformer-base trained with a masked-language-modeling objective · method: CCA / mutual-information layer-trajectory analysis

LLaVA

Visual Representation Alignment for Multimodal Large Language Models (2025)measured

MLLM visual tokens diverge from frozen VFM features; alignment loss recovers accuracy

Details

Yoon et al. measure CKNNA similarity between LLaVA-1.5-7B/13B's internal visual-token hidden states and frozen vision-foundation-model features (DINOv2, CLIP, SAM, Depth Anything V2, RADIO) across all 32 layers [yoon-etal-2025-viral-visual-representation-alignment] Similarity drops sharply after the early layers and only partially recovers in the middle, so the MLLM's visual processing diverges from the pretrained encoder's geometry with depth [yoon-etal-2025-viral-visual-representation-alignment] Adding an explicit cosine-alignment loss toward the frozen VFM causally improves benchmarks (LLaVA-1.5-7B+DINOv2: MMVP 28.20 to 33.33, POPE 85.70 to 88.32, MME 1650.21 to 1694.52) [yoon-etal-2025-viral-visual-representation-alignment] A permutation-sensitivity test shows the aligned model is more spatially sensitive (patch-shuffling degrades accuracy 13.0% with the loss vs 6.5% without) [yoon-etal-2025-viral-visual-representation-alignment]

models: LLaVA-1.5-7B, LLaVA-1.5-13B · method: CKNNA representational alignment, Representation-alignment loss regularization