acceptodds
Under review as a conference paper at ICLR 2027

Testing the Platonic Representation Hypothesis with Life-Event Models

Abstract

Do models trained on life histories and language learn similar relationships between events? We test the Platonic Representation Hypothesis by comparing (a) models trained on nationwide registry histories with (b) independently trained language models given descriptions of the same events. Across life models of increasing size, we measure local alignment, whether the two representation spaces agree on which events are neighbours, and global alignment, whether they agree on the similarities between all pairs of events. Following the Aristotelian critique, we calibrate these comparisons for chance agreement and layer selection. Local alignment increases at every life-model size step for every language model in the original comparison. Better language performance is also associated with greater local alignment, both in this cohort and in a broader comparison of language models. These local-alignment findings persist after controls for shared wording and coding categories. Global alignment also exceeds chance for every pairing, but it stops increasing with life-model size beyond a few million parameters. The results support the local form of the Platonic Representation Hypothesis in a modality that shares no training data with language.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, representational alignment, Platonic Representation Hypothesis, life-event models, large language models, scaling laws, cross-modal alignment

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CPvmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tests the Platonic Representation Hypothesis by comparing autoregressive models trained on Danish administrative life histories with independently trained language models representing descriptions of the same event codes. Across seven life-model sizes and two language-model cohorts, the authors measure local neighborhood agreement using mKNN and global similarity agreement using RBF CKA, with permutation calibration that repeats the complete layer search. Local alignment increases with life-model size and correlates with language-model performance, whereas global alignment remains above chance but plateaus along the life-model scaling axis. The paper validates these findings through controls for lexical overlap, coding hierarchies, representation dimensionality, random initialization, and co-occurrence baselines, providing evidence for partial representational convergence in an additional modality.

Critical strengths

The comparison provides a valuable test of representational convergence using unusually independent training sources. The life models learn from confidential administrative histories without access to event descriptions, while the language models represent descriptions without access to those histories. This setting reduces the shared-training-data explanation that complicates other cross-modal comparisons. Although it does not eliminate shared human or institutional structure, it provides a meaningful extension of the empirical evidence for the hypothesis. The study carefully addresses several plausible measurement artifacts and alternative explanations. Repeating layer selection within each permutation is particularly valuable, and the lexical, coding-hierarchy, dimensionality, and co-occurrence controls strengthen the local-alignment finding. The distinction between increasing local agreement and plateauing global agreement also makes the conclusion more informative than a single aggregate similarity result. These controls and the appropriately qualified findings are the main reasons for my positive recommendation.

Critical weaknesses

The conceptual contribution is primarily an extension of an existing hypothesis and evaluation framework to a new domain. The observed local-versus-global distinction broadly follows earlier findings, and the experiments establish geometric agreement without demonstrating functional interchangeability or a concrete consequence of that agreement. Moreover, administrative codes and their descriptions reflect shared human categories and institutional processes, some of which may remain beyond the measured controls. Consequently, the evidence supports partial agreement between the tested representations more directly than it supports a broader interpretation of convergence toward a shared model of reality. The scaling results leave important attribution questions unresolved. Language-model capability is strongly confounded with model size: in the HellaSwag cohort, local alignment correlates approximately as strongly with log-parameters as with benchmark performance. The conclusions also depend on the representation: contextual vectors show increasing alignment, whereas static embeddings on the larger code set peak at an intermediate size, and fixed-layer comparisons give weaker associations.

Questions for authors
  • Do the available language-model results provide evidence that performance predicts alignment beyond model size and family? A size-adjusted or within-family analysis, with appropriate uncertainty estimates, would help establish whether capability contributes explanatory information beyond scaling.

  • Given the different trends for contextual vectors, static embeddings, and fixed-layer comparisons, what is the intended scope of the convergence claim? Is it specifically about the existence of an increasingly aligned representation somewhere within each model, rather than convergence of a consistently defined representation?

  • Have you tested training on chronologically shuffled histories while preserving the events within each individual’s history? The Markov baseline already captures substantial alignment, so this comparison would help determine how much of the remaining advantage depends on temporal structure.

Reviewer_C9ssOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, they test the Platonic Representation Hypothesis in a new setup. They compare the representations of LLMs with those of “life models”, transformer models from 10k to 500M parameters that they train. They use data from the private Danish registry, covering 20k lives with events related to diagnoses, occupation, industry, education, income, and age. These events form a gallery with more than 23k codes, which they convert into tokens. On the other side, LLMs (12 from the PRH cohort + 25 from the HellaSwag cohort) read only the textual description of each code and never see registry data.

They measure with mKNN (local, k=10, gallery of 1024 codes) and RBF CKA (global), both with the permutation calibration of Gröger et al. (200 shuffles, repeating the layer search inside each one).

They find that the local alignment is above the null and grows with the size of the life model and with the performance of the LLM. The alignment surpasses the untrained co-occurrence baselines (counts), although by a small margin. For the small life models (10k/100k), the results are below the counts. The global alignment is above chance, but it saturates early, and its relation with the capacity of the LLM depends on the cohort.

Critical strengths
  1. As far as I know, this work is the first one to investigate the convergence of representations between LLMs and life event models, contributing to the timely literature on better understanding the Platonic hypothesis. The separation between both modalities is clear: the life model is trained on event IDs only and never sees the descriptions, and the LLM reads only the descriptions and never sees registry data.

  2. It uses the corrected metrics from Gröger et al., and the calibration is done properly: different permutations of the code-description pairing, with the full layer search repeated inside each permutation. Also, there are extensive controls: other-systems-only neighbours, TF-IDF residuals, hierarchy residuals, and a k-sweep. Finally, the negative results are reported honestly.

Critical weaknesses
  1. In the main analysis (Figs. 1, 3, 4), the metric is mKNN with k=10 on a 1024-code gallery (k/n ≈ 1/100), the coarse regime where only category-level structure is measured (Koepke et al., 2026). In Table 9, results on the 15,933-code gallery are reported, where the size trend survives against text embedders, but it is never connected to the main cohorts. So the headline claim is established at a single zoom level where only coarse agreement is expected to survive.

  2. The global capability trend reported by this paper partially contradicts Gröger et al. (2026), especially in the HellaSwag cohort. Gröger et al. found that the correlation collapses after calibration on images, and this paper uses the same 12 LLMs and metric. The paper acknowledges this in App. A (“Their pooled capability ranking differs from our performance measure and per-life-model correlations”), but attributes the difference to pooling and to the capability proxy without running any controlled comparison.

Questions for authors
  • (W1) Run the main comparison (all life model sizes x 12 or 25 LLMs) with a larger gallery or a smaller k/n, and report whether the size trend and the LLM gap hold.

  • (W2) Run the identical calibrated RBF CKA protocol for the same LLMs against DINOv2 on the same dataset as Gröger et al., to report the trend side by side as a control.

  • Report the domain composition of the gallery used. Also, since there is space left, I would move some results from the Appendix to the main body.

Reviewer_nHKvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper checks whether two models that never share training data still agree on which events are similar. The life-event models are trained on Danish registry data, and the LLMs only see event descriptions. Local agreement improves with size and language skill, but global agreement levels off early. The data is from one country and is hard to access, so reproducing it is tough.

Critical strengths

The modality is new, and registry data is a clever testbed since the two sides share no training text. The layer-search-inside-permutation calibration is careful, and the authors report results that cut against their story (global alignment).

Critical weaknesses
  • The headline rests on few effective units: 12 or 25 LLMs, seven life-model sizes, and about 16 fitted runs.
  • Language performance and parameter count move together, so “better at predicting text” and “bigger” can’t be separated.
  • The 10k model is excluded from the fit post hoc, and its control results are unstable.
  • The global-alignment sign flips across sizes and cohorts, and the story is hard to follow. Also the paper doesn’t clearly explain why
  • Reproducibility might be limited to Statistics Denmark access, which should be stated more prominently.
Questions for authors

Can you give confidence intervals over LLMs, not only over codes?

Could a frequency-matched control rule out input frequency as the driver?

Can you test a second country or population subset, at small scale?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.