Representational Convergence

59 beliefs (59 IN, 0 OUT)

The topic of Representational Convergence encompasses the theoretical, methodological, and empirical landscape surrounding whether independently trained neural networks develop shared geometric structure in their internal representations, and if so, in what form and under what conditions. At its center sits the Platonic Representation Hypothesis (prh-convergence-in-distance-measure-not-accuracy), which posits that convergence manifests specifically in how models measure distance between datapoints rather than in task accuracy, and that this convergence to shared global geometric structure logically implies the weaker Aristotelian Representation Hypothesis about local neighborhood relationships, but not conversely (prh-implies-arh-but-not-converse). The formal proof of PRH, however, is bounded: it holds only when the observation function is a bijective projection of the underlying world, and lossy or stochastic observations break the convergence guarantee (prh-formal-proof-requires-bijective-observations). The practical significance of this research lies in distinguishing genuine learned alignment from architectural artifacts, a distinction with direct consequences for model interpretability, cross-modal transfer, and the interpretation of standard benchmarks that may be measuring the same geometric quantity that governs internal feature organization (evaluation-geometry-convergence).

The empirical findings in this cluster converge on a striking local-global dissociation. Before calibration, global spectral metrics such as CKA show strong scaling trends with model capability, but this is largely an artifact: the width confounder drives sample CKA toward 1 as dimensionality grows (cka-width-inflation-gap-exact), and the depth confounder inflates maximum-aggregated scores via multiple comparisons over layer pairs (depth-confounder-scales-as-sqrt-log-m). Proper aggregation-aware calibration requires applying a single shared permutation across all layers to preserve the joint dependence structure exploited by the selection operator (aggregation-aware-calibration-requires-same-permutation-across-layers, aggregation-calibration-shared-permutation), and one must calibrate the aggregated statistic rather than per-element scores to preserve super-uniformity (aggregation-calibration-order-matters). After this calibration, the picture inverts: global metrics lose their scaling trend (linear CKA Pearson r drops from 0.86 to 0.45; Unbiased CKA from 0.65 to −0.01) while local neighborhood metrics like mutual k-NN retain significant alignment (post-calibration-cka-correlation-drops-while-mknn-persists, local-global-dissociation-after-calibration). This pattern holds across text-image, text-video, and text-audio modality pairs (prh-findings-generalize-across-three-modality-pairs), and the qualitative pattern is invariant across significance levels (alpha-invariance-of-local-global-pattern). The cross-modal convergence that does survive is coarse and categorical rather than fine-grained: alignment at k = n/100 remains stable across gallery sizes, while alignment at fixed small k drops dramatically (coarse-grained-cross-modal-convergence-stable-at-k-n100, mutual-knn-captures-coarse-categorical-not-fine-grained-alignment). Models agree on which points are neighbors (topological structure) but not on how close they are (metric structure) (topological-not-metric-cross-modal-alignment). Several additional dissociations sharpen the picture: CKA is higher on problems models collectively fail than on problems they solve (difficulty-inversion-cka-values, difficulty-inversion-cka-higher-on-failed-problems), representational similarity drops sharply from pre-decision to post-decision layers indicating convergence is concentrated in input encoding rather than output generation (generation-gap-pre-vs-post-decision-cka), different-answer problem pairs yield higher CKA than same-answer pairs dissociating representational similarity from output agreement (different-answer-pairs-yield-higher-cka-than-same-answer), and randomly initialized models exhibit higher CKA than trained models, revealing a large architectural component (causal masking, residual streams, layer normalization) beneath the learned signal (random-init-cka-exceeds-trained-cka). The Usama and Chang (2026) study across 16 models and 800 problems formalizes three such dissociations between representational and reasoning convergence (convergence-study-tested-16-models-800-problems), with effects stronger in base models than instruction-tuned variants (convergence-effects-stronger-in-base-models-than-instruction-tuned), and the difficulty inversion effect specifically reversed on GSM8K mathematics problems, attributed to algorithmic compressibility constraining valid solution strategies (gsm8k-reverses-difficulty-inversion).

On the question of belief status, all beliefs in this topic are currently IN; no beliefs have been retracted. The cluster contains a mix of foundational premises and derived findings. Premise-level beliefs include the formal metric definitions (prh-cka-defined-as-normalized-hsic, prh-mnn-per-sample-average-intersection-formula, prh-kernel-formal-definition, rv-coefficient-equals-linear-cka-on-centered-data, procrustes-rotation-invariant-scaling-sensitive), the statistical theorems bounding behavior (bernstein-von-mises-gaussian-convergence, freedman-1965-bvm-countable-failure, width-confounder-cka-null-baseline-o-d-over-n, unbiased-cka-can-be-negative), and the calibration methodology requirements (aggregation-calibration-order-matters, aggregation-calibration-shared-permutation). Derived or empirical findings include the local-global dissociation results, the difficulty inversion, the generation gap, the random-init comparison, and the cross-modal gallery-scaling results. A notable internal tension is captured by the finding that CKA is sensitive to network scale and can be null-calibrated away, while mutual kNN is more stable for local structural comparison, though mutual kNN was previously validated only under small-scale and bijective conditions (cka-sensitive-to-scale-unlike-mutual-knn), and that the mKNN trend peaks at k=10 and flattens by k=100 while CKA-RBF trend depends on bandwidth choice (mKNN-trend-peaks-at-k-10). The geometric framing extends beyond pure representation comparison: the industry trajectory of expanding context windows and agentic loops is read as operationalizing a two-channel architecture of superposed parametric memory and unbounded contextual complement that jointly span a broader effective semantic space (context-trend-as-geometric-convergence-in-practice-v2), while the formal optimality of contrastive losses ties the learning objective back to the same PMI structure that the convergence metrics quantify (prh-infonce-optimal-equals-kpmi-plus-input-dependent-offset, prh-binary-nce-optimal-equals-kpmi-plus-constant). The practical accessibility of these convergent capabilities remains constrained by engineering effort, safety concerns, and licensing tensions that operate orthogonally to raw model capability (frontier-accessibility-gap-persists-despite-capability-convergence-v2).

IN aggregation-aware-calibration-for-depth-confounder
Aggregation-aware calibration computes the null distribution of the same aggregate statistic (e.g., max over all L_A × L_B layer pairs) that is reported, rather than calibrating individual layer-pair scores, to address the depth confounder's multiple-comparisons inflation.
IN aggregation-aware-calibration-requires-same-permutation-across-layers
Aggregation-aware null-calibration requires applying the same row permutation π_k to all layers of one model (S^{(k)}_{ℓ,ℓ'} = s(X_ℓ^A, π_k(Y_{ℓ'}^B))) to preserve the joint dependence structure exploited by the selection operator, rather than calibrating each layer independently and then aggregating.
IN aggregation-calibration-order-matters
Calibrating per-element similarity scores and then aggregating (e.g., taking max) is statistically invalid; one must calibrate the aggregated statistic T(S) against the permutation null {T(S^(k))} to preserve super-uniformity (Proposition C.4)
IN aggregation-calibration-shared-permutation
In aggregation-aware null calibration, a single permutation π must be applied across all layer pairs simultaneously; using independent permutations per layer destroys the dependence structure and invalidates the EVT correction
IN alpha-invariance-of-local-global-pattern
The qualitative pattern of no convergence trend for global metrics and retained trend for local metrics is invariant across significance levels α ∈ {0.01, 0.05, 0.10}; calibration is the critical factor, not the specific σ or α value
IN bernstein-von-mises-gaussian-convergence
The Bernstein-von Mises theorem states that under regularity conditions, the posterior distribution converges to a Gaussian as n→∞, independent of the chosen prior.
IN cka-sensitive-to-scale-unlike-mutual-knn
Central Kernel Alignment (CKA) is sensitive to network scale and can be null-calibrated away (Gröger et al.), while the mutual kNN metric is more stable for local structural comparison, though mutual kNN was previously validated only under small-scale (≤1,024 samples) and bijective evaluation conditions (Huh et al. 2024).
IN cka-width-inflation-gap-exact
For fixed sample size n, the sample centered linear CKA converges to 1 almost surely as width d→∞, with the inflation gap equal to exactly 1−ρ² where ρ² is the true population alignment
IN coarse-grained-cross-modal-convergence-stable-at-k-n100
At k = n/100 (coarse, semantic-category level), cross-modal mutual kNN alignment remains stable as gallery density increases, but at fixed small k (k=1, k=10) alignment drops—indicating vision and language models share broad semantic categories but do not achieve fine-grained representational convergence.
IN context-trend-as-geometric-convergence-in-practice-v2
The industry trajectory of expanding context windows (200K tokens), agentic tooling, and multi-hour autonomous runs operationalizes the parametric/contextual duality as a genuine two-channel architecture: the context window serves as an unbounded, interference-free complement to the finite superposed parametric memory, and iterative agentic loops extend the effective write channel beyond any single forward pass. This product-level trajectory is consistent with the geometric framing that the two channels jointly span a broader effective semantic space than either in isolation, while preserving the architectural distinction between superposed and non-superposed knowledge storage.
IN convergence-effects-stronger-in-base-models-than-instruction-tuned
Difficulty inversion, generation gap, and epiphenomenal correctness effects are stronger in base (pre-trained) models than in instruction-tuned variants, implying fine-tuning partially mitigates but does not eliminate the dissociation.
IN convergence-inversion-persists-at-70b-open-question-at-400b
The difficulty inversion effect is measured at +0.062 at 70B parameters; behavior at 400B+ models (e.g., Nemotron-4 340B) is flagged as an open question.
IN convergence-robustness-thresholds-cka-variance-prompt-sensitivity
The convergence study reports CKA variance < 0.008 and prompt sensitivity < 4% as robustness thresholds (Appendix B).
IN convergence-study-hardware-bfloat16-greedy-decoding
The convergence study used a single NVIDIA RTX 5090 (32 GB VRAM) for most models, 2× A100 80 GB for 70B models, bfloat16 precision, and greedy decoding (temperature = 0).
IN convergence-study-nrf-korea-funding-grant
The convergence study (Usama & Chang, 2026) was supported by the National Research Foundation of Korea (NRF), grant RS-2026-25473622 (MSIT).
IN convergence-study-tested-16-models-800-problems
The Usama & Chang (2026) convergence study tested 16 models (1.5B–72B parameters) on 800 reasoning problems to identify three dissociations between representational and reasoning convergence.
IN depth-confounder-scales-as-sqrt-log-m
The depth confounder from max-aggregation over M = L_A·L_B layer pairs inflates the expected maximum under H₀ as E[T_max] ≤ μ + Cσ√(log M), meaning deeper models receive systematically higher raw alignment scores purely from a larger search space.
IN different-answer-pairs-yield-higher-cka-than-same-answer
Different-answer problem pairs yield higher CKA (0.951) than same-answer pairs (0.906) in 83 out of 91 model pairs, dissociating representational similarity from output agreement.
IN difficulty-inversion-cka-higher-on-failed-problems
Across 14 LLMs, mean pairwise CKA is higher on problems models collectively fail (0.897, 0-4 correct) than on problems they solve (0.830, 10-14 correct), a gap of +0.067 significant at p < 0.001 via 10,000-iteration permutation testing with Benjamini-Hochberg correction (q=0.05).
IN difficulty-inversion-cka-values
Models converge more on problems they collectively fail (CKA = 0.897) than on problems they solve (CKA = 0.830), a phenomenon termed difficulty inversion across 16 LLMs from 8 families (1.5B–72B parameters) on 800 reasoning problems
IN evaluation-geometry-convergence
The convergence of embedding evaluation (cosine → Spearman correlation in SBERT/MTEB) and internal geometry analysis (covariance whitening in ROME, polytope geometry in Park) on the same second-moment structure reveals that standard benchmark evaluation is measuring the same geometric quantity that governs internal feature organization—evaluation and analysis are two readouts of one space.
IN freedman-1965-bvm-countable-failure
Freedman (1965) showed that the Bernstein-von Mises convergence guarantee can fail for infinite countable probability spaces.
IN frontier-accessibility-gap-persists-despite-capability-convergence-v2
Frontier competition drives capability advancement across both proprietary and open-weight models, while safety concerns and licensing/definitional tensions operate as orthogonal constraints on what users can actually access—neither alone determining availability, and resolving one not resolving the other. Given that realizing frontier capabilities in production involves substantial engineering effort beyond model training alone, these combined factors suggest a gap between frontier-level capabilities and widely accessible deployments.
IN generation-gap-pre-vs-post-decision-cka
Representational similarity (CKA) drops from 0.875 in pre-decision layers to 0.274 in post-decision layers (gap = 0.601), with 89 of 91 model pairs exceeding a 0.40 pre-post gap, indicating convergence is concentrated in input-encoding rather than output-generation stages.
IN gsm8k-reverses-difficulty-inversion
GSM8K (mathematics) reverses the difficulty inversion pattern: hard-problem CKA = 0.76 and easy-problem CKA = 0.89 (gap −0.122), unlike science, truthfulness, and commonsense domains which show the standard inversion, attributed to algorithmic compressibility constraining valid solution strategies.
IN laplacian-eigenmaps-laplace-beltrami-convergence
Laplacian Eigenmaps uses eigenfunctions of the Laplace–Beltrami operator as embedding dimensions, and the graph Laplacian converges to the Laplace–Beltrami operator as the number of samples N approaches infinity under mild conditions.
IN layernorm-applied-per-vector-not-batch-dimension
Layer normalization (LayerNorm) is applied element-wise across the feature dimension of a single token vector, not across the batch dimension, and is essential for numerical stability and convergence despite being conceptually unnecessary
IN local-global-dissociation-after-calibration
After permutation calibration, local neighborhood metrics (mKNN, CKNNA, cycle-kNN) retain significant cross-modal alignment and scaling trend with model capability, while global spectral metrics (CKA linear, CKA RBF, SVCCA, RV, Procrustes) lose their scaling trend (CKA linear Pearson r drops from 0.86 to 0.45; Unbiased CKA from 0.65 to −0.01)
IN mKNN-trend-peaks-at-k-10
The calibrated mKNN scaling trend with model capability is strongest at k=10 and flattens by k=100, while CKA-RBF trend emerges at σ=0.5–2.0 but σ=0.1 yields no calibrated alignment
IN mutual-knn-captures-coarse-categorical-not-fine-grained-alignment
Mutual kNN alignment at k = n/100 remains stable across gallery sizes (WIT-1K through LAION-15M), indicating shared coarse semantic-category structure from overlapping web data, while alignment at fixed small k (1, 10) drops dramatically, indicating absence of fine-grained representational convergence between independently trained unimodal encoders.
IN post-calibration-cka-correlation-drops-while-mknn-persists
After null-calibration, the Pearson correlation with model capability drops for global spectral metrics (linear CKA: 0.86→0.45; Procrustes: 0.89→0.39) while local neighborhood metrics remain essentially unchanged (mKNN: ~0.85; CKNNA: ~0.87), indicating global spectral convergence was largely an artifact of width and depth confounders.
IN post-calibration-three-way-distinction-global-local-knn-distances
After permutation-based null-calibration, global spectral measures (CKA, CCA) lose their scaling convergence trend, local neighborhood similarity (mutual k-NN) retains significant cross-modal alignment, but local pairwise distances do not retain alignment.
IN prh-binary-nce-optimal-equals-kpmi-plus-constant
The Bayes optimal solution of the Binary NCE contrastive loss equals K_PMI(xa, xb) + c_X where c_X = log(p_pos/(1−p_pos)) is a global constant, and K_PMI(xa, xb) = log[P_coor(xa,xb) / (P_xa(xa)·P_xb(xb))].
IN prh-caption-density-improves-vision-language-alignment
Using the Densely-Captioned-Images dataset with caption lengths of 5, 10, 20, and 30 words (summarized with LLaMA3-8B-Instruct), higher caption density yields better mutual nearest-neighbor vision-language alignment scores, supporting the information-ceiling sub-hypothesis that more bijective mappings improve convergence.
IN prh-cka-defined-as-normalized-hsic
CKA(K, L) = HSIC(K, L) / √(HSIC(K, K)·HSIC(L, L)) where HSIC(K, L) = (1/(n−1)²)·Trace(K̄·L̄); CKA is global, ordinal, invariant to isotropic scaling, and ranges in [0, 1].
IN prh-cknna-interpolates-between-cka-and-mnn
CKNNA (Centered Kernel Nearest-Neighbor Alignment) recovers CKA as k → dim(K) (full batch) and reduces to mutual k-NN when all nearby samples are treated as equidistant, forming a continuum between the two metrics.
IN prh-code-repository-location
The Platonic Representation Hypothesis project code is hosted at github.com/minyoungg/platonic-rep and the project page is at phillipi.github.io/prh.
IN prh-convergence-in-distance-measure-not-accuracy
The Platonic Representation Hypothesis posits that neural network representations converge specifically in how models measure distance between datapoints as they scale, not merely in task accuracy, across architectures, objectives, and data modalities
IN prh-cross-modal-alignment-predicts-downstream-performance
Higher alignment to vision model DINOv2 correlates linearly with Hellaswag (commonsense reasoning) scores in LLMs and shows an emergence-like trend for GSM8K (math) scores, linking representational convergence to downstream task performance.
IN prh-dataset-pools-wit-1m-laion-15m
The PRH evaluation datasets are constructed as WIT-1M (2,389,146 final pool → 1M sampled) and LAION-15M (17,298,107 final pool → 15M sampled), with the query set fixed at WIT-1024 and gallery scaled to 1K, 10K, 50K, 100K, 500K, 1M (WIT), and 10M, 15M (LAION-400M).
IN prh-early-cnn-layers-converge-first
Early CNN layers (Gabor-like filters) converge across architectures and datasets before later layers do; later layers remain more task-specific, as shown by model stitching results from Lenc & Vedaldi (2015).
IN prh-evaluation-uses-faiss-indexflatl2-and-pHash-dedup
The Koepke et al. (2026) cross-modal alignment evaluation uses Faiss IndexFlatL2 for nearest-neighbor search (equivalent to cosine similarity on L2-normalized vectors), extracts DINOv2 CLS tokens from all layers (DINOv2-B = 768d, G = 1536d) and LLM hidden states from all layers with average pooling (OpenLlama-3B = 3200d), and applies pHash deduplication (64-bit fingerprint, Hamming distance ≤ 2) at large gallery scales.
IN prh-findings-generalize-across-three-modality-pairs
The degradation of cross-modal alignment with gallery size at fixed small k, and the weak/flat LLM-capability-to-alignment trend, are observed consistently for text–image (DINOv2 + OpenLlama), text–video (VideoMAE-v2 on PVD-100k), and text–audio (Dasheng on LAION-Audio-100k), indicating the pattern is a structural property of independent training rather than a text-image-specific artifact.
IN prh-formal-proof-requires-bijective-observations
The formal proof of the Platonic Representation Hypothesis in Section 4 holds only when the observation function is a bijective projection of the underlying world Z; lossy or stochastic observations (e.g., an image cannot convey 'I believe in freedom of speech') violate the assumption and break the convergence guarantee.
IN prh-four-pillars-structure
The Platonic Representation Hypothesis paper is structured around four pillars: (1) literature survey of convergence, (2) empirical demonstration across modalities, (3) identification of select pressures, and (4) discussion of implications, limitations, and counterexamples
IN prh-implies-arh-but-not-converse
The Platonic Representation Hypothesis (convergence to shared global geometric structure) logically implies the Aristotelian Representation Hypothesis (convergence to shared local neighborhood relationships), but the converse does not hold, making ARH a strictly weaker claim.
IN prh-infonce-optimal-equals-kpmi-plus-input-dependent-offset
The Bayes optimal solution of InfoNCE with temperature τ recovers K_PMI plus an input-dependent offset c_X(x_a) (for τ=1) or an additional scale factor (for general τ), and negative pairs are sampled i.i.d. from the marginals P̃(x) = ∫ P_coor(x, x⁺) dx⁺ rather than from the joint distribution.
IN prh-kernel-formal-definition
In the Platonic Representation Hypothesis (Huh et al., ICML 2024), the kernel is formally defined as K(xᵢ, xⱼ) = ⟨f(xᵢ), f(xⱼ)⟩, capturing the similarity/distance structure induced by a representation f: X → ℝⁿ, and it is the kernel (not raw embeddings) that is compared across models.
IN prh-low-k-metrics-reveal-stronger-alignment-than-high-k
Low-k (local) nearest-neighbor metrics (e.g., k=10) reveal a coherent, pronounced alignment trend across model scales and tasks, while high-k (≈ batch size) global metrics show only weak alignment — a counterintuitive result favoring local over global comparison.
IN prh-mnn-alignment-metric-definition
The mutual nearest-neighbor (MNN) alignment metric used in the Platonic Representation Hypothesis paper is defined as the mean intersection of k-NN sets induced by two kernels, normalized by k, and achieves approximately 0.16 in cross-model experiments on a scale where the maximum is 1.0.
IN prh-mnn-cross-model-alignment-score-0-16
Huh et al. (ICML 2024) report a mutual k-nearest neighbor alignment score of 0.16 (theoretical maximum 1.0) when comparing independently trained models, explicitly flagged as an open question rather than a definitive conclusion about the degree of representational convergence.
IN prh-mnn-per-sample-average-intersection-formula
The mNN metric computes per-sample alignment as (1/k)·|S(ϕᵢ) ∩ S(ψᵢ)| where S(ϕᵢ) = d_knn(ϕᵢ, Φ \ {ϕᵢ}) (self-excluded), then averages over the mini-batch; it is local, symmetric, and non-ordinal (neighbor ordering is irrelevant).
IN prh-published-icml2024-mit-four-authors
The Platonic Representation Hypothesis was published at the 41st ICML (Vienna, Austria), PMLR 235, 2024, by Huh, Cheung, Wang, and Isola (all MIT, equal contribution), with preprint arXiv:2405.07987v5 [cs.LG]
IN procrustes-rotation-invariant-scaling-sensitive
Procrustes similarity is invariant to rotation and reflection (orthogonal group) but sensitive to anisotropic scaling, unlike CKA which is invariant to isotropic scaling
IN random-init-cka-exceeds-trained-cka
Randomly initialized models exhibit CKA = 0.864 ± 0.007 while trained models show CKA = 0.612 ± 0.011 (p < 10⁻⁵), indicating a large component of observed cross-model representational convergence is architectural (causal masking, residual streams, layer normalization) rather than learned.
IN rv-coefficient-equals-linear-cka-on-centered-data
The RV coefficient computed on centered representations is algebraically identical to linear CKA (CKA_lin(X,Y) = ‖X_cᵀY_c‖²_F / (‖X_cᵀX_c‖_F · ‖Y_cᵀY_c‖_F)), differing only in historical motivation and in the uncentered case.
IN topological-not-metric-cross-modal-alignment
Cross-modal models agree on which points are neighbors (topological/ordinal structure, shown by mKNN alignment at all k values) but not on how close they are (metric/cardinal structure, shown by CKA-RBF yielding no calibrated alignment at σ=0.1)
IN unbiased-cka-can-be-negative
Unbiased CKA (using the zero-diagonal Gram U-statistic HSIC estimator) can produce negative values at finite n, unlike biased CKA which is always in [0,1]
IN width-confounder-cka-null-baseline-o-d-over-n
Under the null hypothesis, E_H₀[‖Ĉ‖²_F] = d_x·d_y/(n−1), giving a leading-order CKA null baseline of O(d/n) that scales with representation dimensionality relative to sample size.