acceptodds
Under review as a conference paper at ICLR 2027

What Converges in the Platonic Representation Hypothesis? Structure over Geometry

Abstract

The Platonic Representation Hypothesis suggests that increasingly capable models converge toward shared representations. Recent work narrows this claim to shared local neighborhood relationships, finding that capacity-dependent trends in several global similarity measures largely disappear after calibration. We challenge this interpretation by showing that prior local-global comparisons confound structural scale (local versus global) with what is compared: relational structure, defined by which samples are related, versus metric geometry, characterized by quantitative relations such as distances, similarities, or correlations. To disentangle these factors, we construct a controlled framework that evaluates both relational structure and metric geometry at local and global scales. We introduce skeleton overlap as a global counterpart to mutual -nearest neighbors, together with matched distance-aware variants. Across vision-language models, relational structure exhibits robust representational convergence at both scales after calibration, whereas increasingly stringent distance agreement substantially weakens alignment and progressively flattens the capacity-dependent trend. We further extend the analysis beyond ambient Euclidean geometry by evaluating distance agreement under a Riemannian metric approximation and recover the same structure-geometry pattern. The pattern is also reproduced in video-text representations. Together, these results show that relational convergence extends beyond local neighborhoods to global spanning structure, whereas metric geometry exhibits substantially weaker convergence.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Platonic Representation Hypothesis, Representational Convergence, Persistent Homology, Minimum Spanning Trees, Multimodal Representations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_4vUKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper revisits what converges under the Platonic Representation Hypothesis. The authors argue that prior local–global comparisons also change what is compared: local measures typically compare neighbor identities, whereas global measures often involve distances or correlations. This makes it difficult to determine whether the observed differences arise from the scale or the object of comparison.

The authors introduce a 2×2 framework, using nearest-neighbor overlap and minimum spanning tree edge overlap to measure local relations and global spanning structure, respectively. They then introduce distance-agreement requirements while keeping these relations fixed. Image–text and video–text experiments show robust relational alignment at both scales, whereas stricter distance agreement weakens alignment and its association with increasing model size.

Critical strengths
  1. The paper offers a clear and useful refinement of prior conclusions. Separating the scale of comparison from what is compared helps avoid attributing differences between metrics directly to local versus global representational differences. Although the broader topic has been extensively studied, this framework provides a valuable addition to our understanding of representational convergence.

  2. The experimental and statistical procedures are carefully designed. The measures share consistent data, preprocessing, and calibration protocols. Repeating the maximum-over-layer-pairs selection during permutation controls score inflation from layer selection. The authors also examine rank correlations between model size and alignment and preserve dependencies arising from model reuse in their tests. The conclusions therefore go beyond the inevitable score decrease caused by distance weighting.

  3. The main observations receive support from several complementary analyses. The image–text experiments cover 204 model pairs, with additional checks using alternative distance normalizations, cosine similarity, a locally covariance-based metric, and a video–text setting. These results strengthen the credibility of the observed pattern.

Critical weaknesses
  1. It remains unclear whether global skeleton alignment provides evidence beyond local neighborhood agreement. A minimum spanning tree is globally constrained to connect all samples, but its edge overlap may still be dominated by short, local edges. The current results do not sufficiently distinguish these sources. Showing that tree edges outside the local neighborhood relations also exhibit an alignment trend would provide stronger support for the central claim that convergence extends beyond local structure.

  2. Metric stability and statistical power under stringent distance requirements need clarification. The rank-correlation analysis rules out a simple uniform-rescaling explanation, but very strict distance weighting may leave a small number of edges dominating the score. Moreover, the neighborhood and spanning-tree measures involve different numbers of relations, so differences in significance rates may not exclusively reflect differences between local and global geometry. The paper should distinguish geometric disagreement from the effects of sparse effective contributions more clearly.

  3. The interpretation of scaling trends and generalization across settings needs qualification. Model size does not fully represent capability, and training differences remain potential confounders. For example, the VideoMAE scale series mixes pretrained and fine-tuned checkpoints, while some video representations come from single-frame image models. These settings remain informative, but should not be interpreted as isolating model-scale effects or providing comprehensive validation of temporal representations.

Questions for authors
  1. How much of the minimum spanning tree alignment comes from local neighborhood edges? Can existing results or saved edge sets establish whether relations beyond local neighborhoods also contribute to the size-dependent alignment trend? If this has not been analyzed, please clarify the boundary of the claim.

  2. Are the metrics sufficiently stable under stringent distance requirements? Are there existing diagnostics on the number of effectively contributing edges, the prevalence of zero or tied scores, and whether these factors affect the size–alignment correlations or the local–global significance comparison?

  3. How can model-scale effects be distinguished from training differences? In particular, VideoMAE checkpoints at different scales have different training histories. Which cross-modal conclusions remain supported after accounting for such differences?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 4vUK, continued

W1/Q1. Contribution of Local Edges to MST Alignment

In our experiments, all MST edges are contained in the corresponding symmetrized 10-NN graphs (100%). Consequently, no MST edges outside these neighborhoods contribute to the observed overlap or its capacity-dependent trends. We therefore do not claim independent convergence of non-local relations or separately establish between-group convergence.

The local-global distinction in our framework concerns the constraints used to select relations, rather than the lengths of individual edges. While mKNN selects neighbors independently for each sample, MST selects edges under global connectivity and acyclicity constraints. Even locally supported edges can collectively connect the entire representation through multi-edge paths. Therefore, global connectivity does not require individually long-range edges.

Importantly, MST alignment is not mathematically implied by mKNN agreement. To clarify their complementary properties and limitations, we added four one-dimensional representations of six labeled samples in Appendix C.5:

  • X1=(0,1,50,51,100,101)X_1=(0,1,50,51,100,101)
  • X2=(0,50,51,100,101,103)X_2=(0,50,51,100,101,103)
  • X3=(0,1,50,51,101,100)X_3=(0,1,50,51,101,100)
  • X4=(0,1,100,101,50,51)X_4=(0,1,100,101,50,51)
ComparisonStructural difference1-mKNNMST overlap
X1X_1 vs. X2X_2Local neighborhoods change; MST unchanged1/31
X1X_1 vs. X3X_3Rearrangement within a local pair14/5
X1X_1 vs. X4X_4Rearrangement of intact local groups13/5

These examples show that neither measure dominates the other. MST overlap can miss changes in local neighborhoods (X1X_1 vs. X2X_2) and penalize within-group rearrangements that preserve local neighborhood structure (X1X_1 vs. X3X_3). Conversely, MST overlap can detect changes in the relative ordering of local clusters that mKNN misses (X1X_1 vs. X4X_4).

Moreover, setting k=5k=5 makes the neighborhood graphs complete and identical in all comparisons, yielding 5-mKNN =1=1. Nevertheless, MST overlap remains 4/5 for X1X_1 vs. X3X_3 and 3/5 for X1X_1 vs. X4X_4. Thus, even perfect mKNN agreement and complete MST inclusion in the kNN graphs do not guarantee perfect MST overlap. Consequently, convergence in mKNN alone does not establish convergence in MST overlap.

This distinction motivates evaluating MST alignment separately, while not establishing its empirical independence from local neighborhood agreement.

Our main contribution is the controlled 2×22\times2 framework, rather than evidence of independent non-local convergence. We separate relation-selection principles (local neighborhoods versus global spanning constraints) from comparison targets (relational identities versus metric distances). Our findings therefore extend the analysis of relational convergence to globally constrained spanning relations while identifying the distinction between relational structure and metric agreement as the central result.

We revised Sections 3.2 and 4.2 and Appendices C.4–C.5 to clarify this distinction, illustrate the complementary properties of mKNN and MST overlap, and explicitly delimit the scope of our conclusions.

W2/Q2. Metric Stability and Statistical Power Under Stringent Distance Weighting

Stringent distance weighting concentrates alignment scores on a small number of edges, particularly for MST overlap. To quantify this effect, we added diagnostics of effective edge contributions, zero-valued scores, and raw versus calibrated capacity correlations in Appendix E.5.

1. Effective edge contributions and zero-valued scores (Appendix E.5).

The effective number of contributing edges decreases substantially as τ\tau becomes smaller. We quantify this using Neff=(∑ewτ(e))2/∑ewτ(e)2N_{\mathrm{eff}}=(\sum_e w_\tau(e))^2/\sum_e w_\tau(e)^2, which measures the concentration of edge weights.

Across 204 vision–language model pairs, median NeffN_{\mathrm{eff}} for mKNN/MST decreases from 141.6/8.5 at τ=10−2\tau=10^{-2}, to 19.1/2.0 at τ=10−3\tau=10^{-3}, and 3.8/1.0 at τ=10−4\tau=10^{-4}. At the strictest tolerance, all saved raw scores remain nonzero, but null-threshold calibration produces 18 zero-valued mKNN scores and 69 zero-valued MST scores, introducing ties that may affect rank correlations.

2. Effects on capacity-dependent correlations (Appendix E.5).

Capacity correlations weaken even before calibration, showing that calibration-induced ties and uniform score rescaling cannot fully explain the observed changes. From τ=1\tau=1 to 10−410^{-4}, mean within-family Spearman correlations of raw scores decrease:

Scaling directionmKNNMST
Language0.743→0.3330.743\to0.3330.663→0.2650.663\to0.265
Vision0.792→0.6500.792\to0.6500.905→0.5030.905\to0.503

However, sparse effective contributions can also influence raw scores. Therefore, these diagnostics do not establish geometric disagreement as the sole cause of weakening capacity trends.

AuthorsOpenReview ↗

Response to Reviewer 4vUK, continued 2

3. Theoretical interpretation and controlled verification (Appendices E.4 and E.6).

Decreasing τ\tau necessarily reduces the weights of edges with nonzero discrepancies, but does not necessarily weaken the ordering of raw alignment scores across model sizes. To formalize this distinction, we introduce cumulative agreement Gs(t)=1C∣{e∈Es:δs(e)≤t}∣G_s(t) = \frac{1}{C}\left\lvert\left\lbrace e \in E_s : \delta_s(e) \leq t \right\rbrace\right\rvert, defined as the normalized number of shared edges whose distance discrepancies are at most tt.

Proposition 3 establishes that if Gs2(t)≥Gs1(t)G_{s_2}(t)\geq G_{s_1}(t) for every tt, then Ss2(τ)≥Ss1(τ)S_{s_2}(\tau)\geq S_{s_1}(\tau) for every τ>0\tau>0, where Ss(τ)S_s(\tau) is the raw distance-aware alignment score. In other words, if geometric agreement improves with model capacity such that cumulative agreement increases pointwise, then the raw alignment score Ss(τ)S_s(\tau) is monotonically nondecreasing with model capacity for every fixed τ>0\tau>0.

Using a fixed DINOv2 ViT-Base model and 1,024 images, we independently applied Gaussian perturbations in feature space and image space, approaching the same clean reference as noise decreased (Figures 10–11). In both experiments, Gσ(t)G_\sigma(t) preserves its pointwise ordering across noise levels for both mKNN and MST relations, with no violations across the empirical discrepancy breakpoints. Consistent with Proposition 3, raw distance-aware alignment remains monotonic as noise decreases at all five evaluated tolerances, including τ=10−4\tau=10^{-4}.

These results show that strict weighting does not inherently disrupt alignment ordering when representations progressively approach a common reference geometry.

4. Local-global significance comparison (Appendix E.5).

Differences in significance rates cannot be attributed exclusively to differences between local and global geometric agreement. At τ=10−4\tau=10^{-4}, BH-significant pair counts are 183/204 for mKNN and 82/204 for MST.

The two measures involve different numbers of relations and substantially different effective contributions, potentially affecting statistical power. Since these effects have not been independently quantified, we do not interpret the significance gap alone as evidence of weaker global geometric convergence.

We revised Appendices E.4–E.6 to clarify the theoretical distinction, report stability diagnostics, and identify the limitations of interpreting alignment under extreme distance tolerances.

W3/Q3. Scaling Trends and Generalization

The structure-geometry pattern persists within separate pretrained (PT) and fine-tuned (FT) VideoMAE scale series. Since the original series mixed pretrained Base/Large and fine-tuned Huge checkpoints, we conducted an additional 3×2 VideoMAE experiment covering Base, Large, and Huge at both training stages. Each checkpoint was evaluated against the same 13 language models (78 pairs), using aggregation-aware calibration with 500 permutations.

Mean calibrated relational alignment across 13 language models:

MeasureStageBaseLargeHuge
mKNNPT0.02720.04410.0507
FT0.07560.10650.1343
H0H_0 skeleton overlapPT0.01550.02590.0290
FT0.03580.05320.0706

For both measures and both training stages, calibrated alignment increases strictly from Base to Large to Huge for all 13 individual language models, with all 78 pairs significant after BH-FDR correction.

Under stricter distance agreement, these trends become less consistent. At τ=10−2\tau=10^{-2}, strict increases remain for 4/13 (mKNN) and 3/13 (MST overlap) language models in PT, versus 13/13 and 8/13 in FT. At τ=10−4\tau=10^{-4}, only 35/78 distance-aware mKNN pairs and 0/78 distance-aware MST overlap pairs remain significant, although differences in the number of relations may affect statistical power.

These results show that the structure-geometry pattern persists within separate PT and FT scale series, rather than depending solely on their mixture. Nevertheless, other training differences remain uncontrolled, so we interpret these findings as model-size associations rather than causal scaling effects. Model size should not be equated with capability.

Finally, this additional analysis uses VideoMAE exclusively, while DINOv2 and CLIP serve as single-frame baselines in the original evaluation. Our conclusions therefore support cross-modal relational alignment, but do not establish comprehensive convergence of temporal representations.

We have incorporated these findings in Section 4.5 and Appendix H.3.

Thank you for your detailed rebuttal. As someone who also reviews and responds to rebuttals, I found your response helpful in resolving my questions. I appreciate your efforts in addressing the concerns. Thank you again for your rebuttal.

We are glad that our rebuttal has addressed your concerns. Your questions helped us clarify the interpretation of MST overlap and stringent distance tolerances. The additional VideoMAE experiments allowed us to better distinguish model-size trends from training-stage differences. We sincerely appreciate your thoughtful feedback, which has greatly helped us improve our paper.

Reviewer_PWM2OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Here, the authors study the question of the Platonic representation hypothesis (PRH) and the Aristotelian representation hypothesis (ARH) and extend the evaluation currently only done on local neighborhoods to global structures.

Critical strengths
  • The authors have investigated a critical gap in current representational similarity metrics: there were no metrics considering the global structure of representations.
  • The new findings also can have broad implications for alignment of unimodal models (the typical use case of the PRH/ARH).
Critical weaknesses
  • W1. The main weakness of the paper is that the framing the authors adopt throughout is stronger than the results or prior work support. The paper repeatedly claims to challenge the local neighborhood interpretation of the ARH, but the results extend the analysis to another relational structure. They do not show that the previous empirical findings were incorrect, which they make sound throughout the paper. The current framing misleads the results that are presented and also the community. For me, the paper is a significant contribution even without this "refuting result framing" and thus would benefit from presenting the new metrics and the insights as an extension and being more precise about what changes in the interpretation.
  • W2. For me, it is unclear how much of the MST overlap comes from relations beyond local neighborhoods. Especially since an MST connects the entire sample set, but many of its edges can still connect nearby samples. Since the score counts all edges equally, it could be high even if the edges connecting different regions disagree. Appendix C.4 explains why an MST can contain edges absent from a kNN graph, but the authors do not show whether those edges account for the observed trend.
  • W3. While reading the paper, I felt that the choice of weighting in Eq. 3 and the tolerance range would benefit from more motivation. Decreasing τ\tau reduces the weight of every edge with a nonzero discrepancy. Consider at τ=10−4\tau=10^{-4}, normalized distances need to agree to approximately 0.01% to receive a weight of at least e−1e^{-1}. So for me it was unclear why this precision is meaningful here. A more intuitive experiment would have been a controlled experiment with shared geometry and varying noise, which would help distinguish geometric disagreement from instability under very strict weighting.
  • W4. I find the Riemannian extension interesting, but it needs a more complete analysis. This analysis itself would be very novel in itself but currently the paper uses it mainly as an alternative formulation of distances rather than respecting the new geometric aspects that this would open for investigation.
Questions for authors

See weaknesses.

2 comments

W1. Clarifying Our Contribution as an Extension of Prior Work

We thank the reviewer for this important comment. Our intention was not to refute the empirical findings of prior work. However, we recognize that some of our wording could give this impression. We therefore revised the manuscript to clarify our intended contribution.

Our study builds directly on Gröger et al. (2026). They showed that model scale and layer-wise aggregation can inflate similarity scores. After calibration, several global similarity measures lose their capacity-dependent trends, while local neighborhood relationships retain robust convergence. They also found that corresponding pairwise distances do not exhibit comparable convergence. We accept these findings and adopt their aggregation-aware permutation calibration.

Our contribution is to extend this analysis along two dimensions: structural scale (local versus global) and what is compared (relational structure versus metric geometry). This extends the scope of the analysis without contradicting previous empirical findings.

To avoid misunderstanding, we revised the Abstract, Introduction, Contributions, and Conclusion. The revised Introduction now explicitly builds on prior findings and examines whether relational convergence also occurs under global spanning constraints. We appreciate the reviewer's suggestion, which helped us communicate our contribution more precisely.

W2. What MST Overlap Captures Beyond Local Neighborhood Agreement

In our experiments, all MST edges are contained in the corresponding symmetrized 10-NN graphs. Thus, the observed MST alignment is supported by local edges, and we do not claim independent convergence of non-local or between-group relations. We acknowledge that equal-weighted MST overlap can remain high even when a small number of inter-region connections disagree.

The local-global distinction in our framework concerns the constraints used to select relations, rather than the lengths of individual edges. While mKNN selects neighbors independently for each sample, MST selects edges under global connectivity and acyclicity constraints. Even locally supported edges can collectively connect the entire representation through multi-edge paths. Therefore, global connectivity does not require individually long-range edges.

Importantly, MST alignment is not mathematically implied by mKNN agreement. To clarify their complementary properties and limitations, we added four one-dimensional representations of six labeled samples in Appendix C.5:

  • X1=(0,1,50,51,100,101)X_1=(0,1,50,51,100,101)
  • X2=(0,50,51,100,101,103)X_2=(0,50,51,100,101,103)
  • X3=(0,1,50,51,101,100)X_3=(0,1,50,51,101,100)
  • X4=(0,1,100,101,50,51)X_4=(0,1,100,101,50,51)
ComparisonStructural difference1-mKNNMST overlap
X1X_1 vs. X2X_2Local neighborhoods change; MST unchanged1/31
X1X_1 vs. X3X_3Rearrangement within a local pair14/5
X1X_1 vs. X4X_4Rearrangement of intact local groups13/5

These examples show that neither measure dominates the other. MST overlap can miss changes in local neighborhoods (X1X_1 vs. X2X_2) and penalize within-group rearrangements that preserve local neighborhood structure (X1X_1 vs. X3X_3). Conversely, MST overlap can detect changes in the relative ordering of local clusters that mKNN misses (X1X_1 vs. X4X_4).

Moreover, setting k=5k=5 makes the neighborhood graphs complete and identical in all comparisons, yielding 5-mKNN =1=1. Nevertheless, MST overlap remains 4/5 for X1X_1 vs. X3X_3 and 3/5 for X1X_1 vs. X4X_4. Thus, even perfect mKNN agreement and complete MST inclusion in the kNN graphs do not guarantee perfect MST overlap. Consequently, convergence in mKNN alone does not establish convergence in MST overlap.

This distinction motivates evaluating MST alignment separately, while not establishing its empirical independence from local neighborhood agreement.

Our main contribution is the controlled 2×22\times2 framework, rather than evidence of independent non-local convergence. We separate relation-selection principles (local neighborhoods versus global spanning constraints) from comparison targets (relational identities versus metric distances). Our findings therefore extend the analysis of relational convergence to globally constrained spanning relations while identifying the distinction between relational structure and metric agreement as the central result.

We revised Sections 3.2 and 4.2 and Appendices C.4–C.5 to clarify this distinction, illustrate the complementary properties of mKNN and MST overlap, and explicitly delimit the scope of our conclusions.

W3. Motivation for Distance Weighting and Tolerance Selection

1. Motivation for distance weighting and tolerance selection.

Our tolerance sweep examines whether capacity-dependent relational alignment persists under increasingly stringent metric agreement. For each shared edge ee, we use the exponential weight wτ(e)=exp⁡(−δ(e)/τ)w_\tau(e)=\exp(-\delta(e)/\tau), where δ(e)\delta(e) denotes the absolute log-distance discrepancy between the two representations after distance normalization, and τ>0\tau>0 controls tolerance to this discrepancy.

At τ=10−4\tau=10^{-4}, obtaining a weight of at least e−1e^{-1} requires approximately 0.01% relative agreement between normalized distances. However, e−1e^{-1} is not a required weight threshold: smaller weights still contribute and we adjust the y-axis scale to visualize capacity-dependent trends. We do not assign intrinsic geometric significance to this precision; τ=10−4\tau=10^{-4} serves as an extreme sensitivity setting rather than a required precision threshold.

2. Theoretical justification (Appendix E.4).

Decreasing τ\tau necessarily reduces the weights of edges with nonzero discrepancies, but does not necessarily weaken the ordering of raw alignment scores across model sizes. We introduce cumulative agreement Gs(t)=1C∣{e∈Es:δs(e)≤t}∣G_s(t) = \frac{1}{C}\left\lvert\left\lbrace e \in E_s : \delta_s(e) \leq t \right\rbrace\right\rvert , defined as the normalized number of shared edges whose distance discrepancies are at most tt.

Proposition 3 establishes that if Gs2(t)≥Gs1(t)G_{s_2}(t)\geq G_{s_1}(t) for every tt, then Ss2(τ)≥Ss1(τ)S_{s_2}(\tau)\geq S_{s_1}(\tau) for every τ>0\tau>0, where Ss(τ)S_s(\tau) is the raw distance-aware alignment score. Thus, pointwise improvement in geometric agreement preserves raw capacity ordering at every tolerance.

3. Controlled noise experiments (Appendix E.6).

Following the reviewer's suggestion, we conducted two controlled noise experiments and found that stringent weighting preserves alignment ordering under progressively decreasing noise. Using a fixed DINOv2 ViT-Base model and 1,024 images, we independently applied Gaussian perturbations in feature space and image space, approaching the same clean reference as noise decreased (Figures 10–11).

In both experiments, Gσ(t)G_\sigma(t) preserves its pointwise ordering across noise levels for both mKNN and MST relations, with no violations across the empirical discrepancy breakpoints. Consistent with Proposition 3, raw distance-aware alignment remains monotonic as noise decreases at all five evaluated tolerances, including τ=10−4\tau=10^{-4}.

These results show that strict weighting does not inherently disrupt alignment ordering when representations progressively approach a common reference geometry.

4. Empirical interpretation and stability (Appendix E.5).

In actual model comparisons, tolerance-dependent trends reflect both the distribution of distance discrepancies and the increasing concentration of edge contributions. Figure 9 illustrates the former: AugReg exhibits crossings in global cumulative agreement, consistent with its non-monotonic global capacity trend at τ=10−2\tau=10^{-2}, whereas CLIP-FT exhibits more consistent local ordering and a more stable local capacity trend.

Our stability diagnostics show that extreme tolerances concentrate effective edge contributions, but capacity correlations also weaken before calibration, indicating that calibration-induced ties alone cannot explain the observed changes.

Thus, noise experiments establish that strict weighting need not disrupt ordering, while sparse contributions limit interpretations at extreme tolerances.

We revised Sections 3.3 and 4.3 and Appendices E.4–E.6 to clarify the weighting motivation, theoretical interpretation, controlled experiments, and stability limitations.

W4. Riemannian extension

Our Riemannian extension should be interpreted as a locally covariance-adapted metric approximation, rather than a recovery of intrinsic manifold geometry. Its purpose is to examine whether the observed structure-geometry pattern persists under an alternative metric, not to provide a comprehensive analysis of Riemannian geometry.

Our construction uses the spherical logarithm map to obtain local tangent vectors and estimates a covariance-adapted metric from their local covariance structure, motivated by Singer and Coifman (2008) and Berry and Sauer (2016). However, our setting of 1,024 samples and 30 nearest neighbors per point may be insufficient for reliably recovering intrinsic geometry in high-dimensional representation spaces.

A comprehensive Riemannian analysis would require systematic investigation of sampling density, neighborhood scales, and geodesic distance estimation, beyond this study's scope.

We revised Section 3.5 and the Conclusion to clarify the approximate nature of our construction, its role as a robustness analysis, and the limitations of interpreting it as intrinsic geometric recovery.

Reviewer_qWXXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Huh et al. who originally presented the Platonic representation theory used mKNN similarity as local similarity measure to show how alignment of representations of different models generally increases as model sizes increases. In the appendix, they state that CKA as global similarity measures did not show this trend as clearly. Groger et al. who present an alternative Aristotle’s representation theory revisit this and calibrate CKA and mKNN similarity to model depth and width. They find that local similarity (mKNN) converges but global (CKA) does not. This work further refines these findings by disentangling (global, local)x(relational structure, metric geometry) similarity. To this end, they propose a new global relational structure similarity called H0H_0 based on shared edges in the MST of sample representations between two models. They find that across vision-language model global relational structure shows representational convergence. They also introduce distance-aware variants of mKNN and H0H_0.

Critical strengths

This work tackles an interesting question and provides a novel disentanglement of structural and geometric representation geometry on local vs global scale. Previous works have only analysed structural geometry on local and geometric similarity on global scale, leading them to the conclusion that representational alignment converges with model size only on the local, not on global scale. The more fine-grained analysis this work provides challenges the result on no convergence on the global scale.

Critical weaknesses

This work limits the analysis to one form of global structure - which is agreement between minimum spanning tree edge sets. This limitation is acknowledged by the authors, but as they contrast global structural H0H_0 alignment with local structural mKNN alignment, it would be good to contrast what either measure can capture and what it cannot capture beyond the connectivity-based distinction given in Appendix C.4.

First, mKNN measures local neighbourhood agreement. A complementary global structural alignment measure that explicitly compares between-group structure might allow for broader conclusions. H0H_0 alignment can mix within-group edges and between-group edges and only compares shared MST edges between two representations in a binary manner, without distinguishing their roles as between-group or within-group connection. One concern is whether agreement in a MST edge set establishes convergence beyond local neighbourhood relationships or if it is dominated by local edges.

Second, related to my point about clarifying what H0H_0 can and cannot capture, here is a toy example to illustrate that H0H_0 cannot capture agreement in cluster organisation. Here, H0H_0 yields perfect alignment for a representation pair with different clusters and reduced H0H_0 for a representation pair with identical clusters.

SampleRepresentation XRepresentation YRepresentation Z
A000
B11011
C101102102
D102203101

X vs. Y: different clusters (if clusters are the connected components formed by joining samples at distance <2) with perfect H0(X,Y)=1H_0(X,Y) = 1 overlap

X vs. Z: identical clusters, but H0(X,Z)=2/3H_0(X,Z)=2/3 (TX=(AB,BC,CD)T_X=(AB,BC,CD) and TZ=(AB,BD,CD)T_Z=(AB,BD,CD))

Questions for authors
  • The purpose of connecting connection the MST edges to death edges from Vietoris-Rips filtration in Proposition 1 is unclear to me. How does this connection help interpret the measure beyond its definition as MST edge overlap?
3 comments

What MST Overlap Captures and Its Limitations

In our experiments, all MST edges are contained in the corresponding symmetrized 10-NN graphs. Thus, the observed MST alignment is supported by local edges, and we do not claim independent convergence of non-local or between-group relations.

The local-global distinction in our framework concerns the constraints used to select relations, rather than the lengths of individual edges. While mKNN selects neighbors independently for each sample, MST selects edges under global connectivity and acyclicity constraints. Even locally supported edges can collectively connect the entire representation through multi-edge paths. Therefore, global connectivity does not require individually long-range edges.

Importantly, MST alignment is not mathematically implied by mKNN agreement. To clarify their complementary properties and limitations, we added four one-dimensional representations of six labeled samples in Appendix C.5:

  • X1=(0,1,50,51,100,101)X_1=(0,1,50,51,100,101)
  • X2=(0,50,51,100,101,103)X_2=(0,50,51,100,101,103)
  • X3=(0,1,50,51,101,100)X_3=(0,1,50,51,101,100)
  • X4=(0,1,100,101,50,51)X_4=(0,1,100,101,50,51)
ComparisonStructural difference1-mKNNMST overlap
X1X_1 vs. X2X_2Local neighborhoods change; MST unchanged1/31
X1X_1 vs. X3X_3Rearrangement within a local pair14/5
X1X_1 vs. X4X_4Rearrangement of intact local groups13/5

These examples show that neither measure dominates the other. MST overlap can miss changes in local neighborhoods (X1X_1 vs. X2X_2) and penalize within-group rearrangements that preserve local neighborhood structure (X1X_1 vs. X3X_3). Conversely, MST overlap can detect changes in the relative ordering of local clusters that mKNN misses (X1X_1 vs. X4X_4).

Moreover, setting k=5k=5 makes the neighborhood graphs complete and identical in all comparisons, yielding 5-mKNN =1=1. Nevertheless, MST overlap remains 4/5 for X1X_1 vs. X3X_3 and 3/5 for X1X_1 vs. X4X_4. Thus, even perfect mKNN agreement and complete MST inclusion in the kNN graphs do not guarantee perfect MST overlap. Consequently, convergence in mKNN alone does not establish convergence in MST overlap.

This distinction motivates evaluating MST alignment separately, while not establishing its empirical independence from local neighborhood agreement.

Our main contribution is the controlled 2×22\times2 framework, rather than evidence of independent non-local convergence. We separate relation-selection principles (local neighborhoods versus global spanning constraints) from comparison targets (relational identities versus metric distances). Our findings therefore extend the analysis of relational convergence to globally constrained spanning relations while identifying the distinction between relational structure and metric agreement as the central result.

We revised Sections 3.2 and 4.2 and Appendices C.4–C.5 to clarify this distinction, illustrate the complementary properties of mKNN and MST overlap, and explicitly delimit the scope of our conclusions.

1. Motivation for Appendix C.5

We particularly appreciate the reviewer's illustrative counterexample, which motivated us to clarify what MST overlap can and cannot capture. In response, we added Appendix C.5 to systematically contrast the complementary properties and limitations of mKNN and MST overlap.

Our examples show that MST overlap can miss changes in local neighborhoods, penalize within-cluster rearrangements that preserve local neighborhood structure, and detect changes in the relative ordering of local clusters that mKNN misses. These examples clarify the distinct information captured by each measure without suggesting that MST overlap fully characterizes cluster organization.

2. Interpretation of Proposition 1

Regarding Proposition 1, its purpose is to provide a topological interpretation of MST edge selection through H0H_0 persistence. Specifically, MST edges correspond to H0H_0 death events, identifying the relations responsible for merging connected components during the Vietoris-Rips filtration.

The broader significance of this connection is its potential for extending the PRH analysis to higher-dimensional topological structures. Our current measure retains only edge identities, not death times or the full cluster hierarchy. However, persistent homology provides a natural framework for investigating whether cycles and voids, captured by HkH_k (k≥1k\geq1), are shared across representations.

Such extensions require new alignment measures and remain beyond the scope of this study. Nevertheless, we believe that investigating which topological features converge across representations is a promising direction for understanding the PRH. We have highlighted this direction in Section 3.2 and the Conclusion.

Thanks for the clarifications, in particular the illustrative examples that show how mKNN and MST complement each other and what MST cannot capture. This helped to resolve my questions. I will maintain my original rating. An alignment measure that captures more of the full cluster hierarchy would also be an interesting next step.

We are glad that our clarifications and examples addressed your questions. Your illustrative counterexample inspired the additional examples and helped us better clarify the complementary properties and limitations of mKNN and MST overlap. We also appreciate your suggestion to explore the full cluster hierarchy in future work. Thank you again for your thoughtful feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.