arXiv paper2026-09-23· machine learning (representation learning)· empirical study

What Converges in the Platonic Representation Hypothesis? Structure over Geometry

Q: What is H0 skeleton overlap in this paper?

It is a new global similarity measure for neural network representations: two representations are compared by the overlap between their minimum spanning tree edge sets (n−1 edges over n samples), normalized by n−1. The paper proves these MST edges are exactly the 'death edges' of zero-dimensional persistent homology under the Vietoris–Rips filtration. Because it checks only which sample pairs form the spanning structure — not their distances — it measures global relational structure, playing the same role for the global scale that mutual k-nearest-neighbors (mKNN) plays locally.

Q: Does the Platonic Representation Hypothesis still hold after calibration?

Partially, according to this paper. After aggregation-aware permutation calibration, models converge in relational structure: both local kNN neighborhoods and global minimum-spanning-tree structure show robust, capacity-dependent alignment across 204 vision-language model pairs (all pairs statistically significant). However, models do not converge in metric geometry — requiring increasingly strict agreement on distances sharply weakens alignment (to about 5–6% of the relational baseline at τ=10⁻²) and significantly flattens the capacity trend. So convergence is in structure, not in precise geometry.

Q: How does this paper differ from Gröger et al.'s 'Aristotelian view' of the Platonic Representation Hypothesis?

Gröger et al. (ICML 2026) found that after calibration, global similarity measures (CKA, SVCCA, Procrustes) lose their scaling trends while local mKNN keeps them, concluding that convergence is essentially local. This paper argues that comparison confounds scale with comparison type: the local measure is relational while the global measures are geometry-dependent. Using a matched 2×2 design with a new global relational measure (H0/MST skeleton overlap), it shows global spanning structure converges just as robustly as local neighborhoods — the real dividing line is relational structure versus metric geometry, not local versus global.

Q: What data and models were used in the experiments?

The main experiments use 1,024 paired image-text samples from the WIT subset of the public PRH benchmark, comparing 12 language models (BLOOMZ, OpenLLaMA, LLaMA families) with 17 vision transformers (ImageNet-21K supervised, MAE, DINOv2, CLIP variants), yielding 204 model pairs. A video-text extension uses 1,024 videos from the PVD test split with VideoMAE and frame-level DINOv2/CLIP (143 pairs). All alignment scores are maximized over layer pairs and calibrated with 500 correspondence permutations; significance uses Benjamini–Hochberg FDR control. The authors state code is available at github.com/junwon0/what-converges-prh.

Junwon You, Mihyun Jang, Sangwoo Mo, Jae-Hun Jung

View original paper ↗arXiv:2609.27252

cs.LG · cs.CV · math.AT

Quick verdict

Argues PRH convergence is in relational structure (local mKNN and global H0/MST overlap), not metric geometry; a controlled, statistically careful reframing.

Summary

The paper revisits the Platonic Representation Hypothesis (PRH) and the recent claim that, after aggregation-aware calibration, only local neighborhood agreement (mKNN) retains capacity-dependent convergence. The authors argue prior local-vs-global comparisons confound structural scale with what is compared: relational structure (which samples are related) versus metric geometry (quantitative distances). They build a 2×2 framework: local mKNN versus global H0 skeleton overlap (overlap of MST edge identities, proven equal to H0 death edges of the Vietoris–Rips filtration), each with distance-aware variants whose weight exp(−|log-distance mismatch|/τ) tightens from τ=1 to 10⁻⁴. On the PRH WIT benchmark (1,024 samples; 12 language × 17 vision models = 204 pairs; 500 permutations; Benjamini–Hochberg FDR), relational structure converges at both scales (all 204 pairs significant; calibrated mKNN and H0 overlap correlate at Spearman ρ=0.968), while stricter distance agreement decays alignment (85.6%→5.8% local and 83.0%→5.5% global of baseline from τ=1 to 10⁻²) and significantly flattens capacity trends (exact permutation tests, all q<0.05). The pattern persists under a Riemannian metric approximation and in video-text representations (143 pairs). Limitations: a single 1,024-sample dataset, one operationalization of global structure, and no quantification of how many MST edges are genuinely non-local.

Review complete

Assessment at a glance

machine learning (representation learning)

risk flags identified

● high 0

● medium 2

● low 1

Reproducibility readinesshigh

Code & data availableyes

Review confidencehigh

AI text signal0/20 passages flagged

Deep research viableCode-reproducibleNeeds GPU

4 strengths · 4 weaknesses · 1 author questions

AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.

Levels reflect verification complexity — not a quality score. Evolve never rates your paper.

Key claims under review

Prior local-vs-global comparisons of calibrated representational convergence confound structural scale with whether relational structure or metric geometry is compared.
Relational structure converges robustly at both local and global scales after calibration: all 204 vision-language pairs remain FDR-significant for mKNN and H0 skeleton overlap, which correlate at Spearman ρ=0.968.
Increasingly stringent distance agreement weakens alignment (85.6%→5.8% local and 83.0%→5.5% global of baseline from τ=1 to 10⁻²) and significantly flattens capacity-dependent trends (all exact permutation tests q<0.05).
The structure-geometry pattern persists under a locally estimated Riemannian metric approximation and in video-text representations (143 pairs).

Strengths

Genuinely controlled 2×2 design
Holds structural support fixed while varying metric agreement (§3.3), so scale and comparison type are varied independently — a real advance over the confounded local-vs-global contrast in Gröger et al. (2026) that the paper targets.
Principled new measure with correctness proof
H0 skeleton overlap is defined as MST edge-identity overlap (Eq. 2); Proposition 1 (App. C.2) proves H0 death edges of the Vietoris–Rips filtration equal MST edges, with deterministic tie handling via Prim (App. C.3) and a formal Prop 2 (App. C.4) showing fixed-k graphs can miss cross-cluster edges.
Statistics that separate substance from mechanics
Extensive robustness and replication

Weaknesses

The 'global' claim is under-quantified
Scale divergence lives only at the extreme τ
Capacity proxy confounded with training paradigm
Small, single-dataset support

Full Review Report

$1.99 · one-time

Unlock the complete preliminary analysis Evolve produced for this paper.

Report preview

Scope & Claims

The paper asks a sharpened version of the Platonic Representation Hypothesis (PRH) question: not whether models converge, but *what* converges. Its target is the recent 'Aristotelian' reinterpretation (Gröger et al., ICML 2026) that after aggregation-aware calibration only local neighborhood agreement (mKNN) retains capacity-dependent scaling, while global measures (CKA, SVCCA, Procrustes) lose it. The central claim: that local-vs-global contrast confounds structural scale with what is compared — relational structure (which samples are related) versus metric geometry (quantitative distances/similarities). Using a controlled 2×2 design, the authors claim (1) relational structure converges robustly at both local and global scales after calibration; (2) increasingly stringent distance agreement weakens alignment and flattens capacity trends at both scales; (3) the pattern is robust to a Riemannian metric approximation and to a video-text modality. This is a conceptual reframing plus an empirical reanalysis of a public benchmark; claims are appropriately scoped, with limitations honestly stated in §5.

Methodology Assessment

The 2×2 framework is the paper's core contribution and is genuinely controlled. Local relational structure: mKNN (k=10). Global relational structure: new H0 skeleton overlap — overlap of minimum-spanning-tree edge identities normalized by n−1 edges. Proposition 1 (App. C.2) correctly proves H0 death edges of the Vietoris–Rips filtration equal MST edges under deterministic tie-breaking (Prim implementation, App. C.3); Prop 2 (App. C.4) gives a constructed worst case where fixed-k graphs miss cross-cluster edges while the MST spans them. Metric geometry is probed by we

The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.

All 4 strengths & 4 weaknesses, analyzed in detail
3 risk flags with severity, section & reviewer questions
Methods & statistical evidence assessment
1 questions a peer reviewer would ask
Full six-dimension evaluation report & recommendations

One-time purchase · this paper only

One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.

Frequently asked questions

What is H0 skeleton overlap in this paper?

It is a new global similarity measure for neural network representations: two representations are compared by the overlap between their minimum spanning tree edge sets (n−1 edges over n samples), normalized by n−1. The paper proves these MST edges are exactly the 'death edges' of zero-dimensional persistent homology under the Vietoris–Rips filtration. Because it checks only which sample pairs form the spanning structure — not their distances — it measures global relational structure, playing the same role for the global scale that mutual k-nearest-neighbors (mKNN) plays locally.

Does the Platonic Representation Hypothesis still hold after calibration?

Partially, according to this paper. After aggregation-aware permutation calibration, models converge in relational structure: both local kNN neighborhoods and global minimum-spanning-tree structure show robust, capacity-dependent alignment across 204 vision-language model pairs (all pairs statistically significant). However, models do not converge in metric geometry — requiring increasingly strict agreement on distances sharply weakens alignment (to about 5–6% of the relational baseline at τ=10⁻²) and significantly flattens the capacity trend. So convergence is in structure, not in precise geometry.

How does this paper differ from Gröger et al.'s 'Aristotelian view' of the Platonic Representation Hypothesis?

Gröger et al. (ICML 2026) found that after calibration, global similarity measures (CKA, SVCCA, Procrustes) lose their scaling trends while local mKNN keeps them, concluding that convergence is essentially local. This paper argues that comparison confounds scale with comparison type: the local measure is relational while the global measures are geometry-dependent. Using a matched 2×2 design with a new global relational measure (H0/MST skeleton overlap), it shows global spanning structure converges just as robustly as local neighborhoods — the real dividing line is relational structure versus metric geometry, not local versus global.

What data and models were used in the experiments?

The main experiments use 1,024 paired image-text samples from the WIT subset of the public PRH benchmark, comparing 12 language models (BLOOMZ, OpenLLaMA, LLaMA families) with 17 vision transformers (ImageNet-21K supervised, MAE, DINOv2, CLIP variants), yielding 204 model pairs. A video-text extension uses 1,024 videos from the PVD test split with VideoMAE and frame-level DINOv2/CLIP (143 pairs). All alignment scores are maximized over layer pairs and calibrated with 500 correspondence permutations; significance uses Benjamini–Hochberg FDR control. The authors state code is available at github.com/junwon0/what-converges-prh.

About this review

This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.

Figures are only partially recoverable from extracted text; visual claims (axis patterns in Figs. 2–12) were assessed from captions and in-text numbers, not from rendered plots.
Equations and some notation were reconstructed from linearized text; proposition statements were cross-checked against surrounding prose.
Some appendix details (exact per-model layer counts) are referenced but not fully enumerated in the extracted text.

Go deeper on this paper

Wrote a paper? Stress-test it before reviewers do.

Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.

Scan your paper