acceptodds
Under review as a conference paper at ICLR 2027

A Theory of Platonic Representations in Language Models

10 views

Abstract

Representations of translated sentences are similar in the inner layers of multilingual language models – an observation connected to the platonic representation hypothesis, yet unexplained theoretically. We provide an explanation based on the assumption that data have a hidden hierarchical structure whose abstract levels are shared across languages while surface levels are modality- or language-specific. Concretely, we generate synthetic languages from probabilistic context-free grammars sharing upper-level but not lower-level production rules. In this setting the Bayes-optimal next-token predictor is belief propagation (BP); encoding its messages in successive layers yields analytical predictions that agree well with transformers trained on the same data. The framework explains why cross-lingual similarity peaks in middle layers, coexists with language-specific structure, and strengthens with language proximity, model quality and data exposure. It distinguishes similarity (shared neighborhood geometry) from alignment (shared coordinates), showing that the latter occurs when code-switched data, i.e. mixed-language sentences, are abundant enough. It further predicts that subtracting from each layer the component linearly predictable from the preceding one increases cross-lingual similarity, which we confirm in pretrained LLMs.

Primary area:
interpretability and explainable AI
Keywords:
platonic representation hypothesis, multilingual LLMs, cross-lingual transfer, context-free grammar, belief propagation

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_QHDjOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper tries to provide a theoretical proof with a single assumption that data possess a hidden hierarchy that is shared at the top (semantic content) and separated at the bottom (structure). The paper experiments with synthetic data modeled with a multilingual random hierarchy model (RHM) and compares a GPT-style decoder-only transformer with a theoretical belief propagation (BP) network. Results showed that geometric similarity was highest on the middle layer, with the emergence of surface morphosyntax layers in the early and later stages of the layers. The same findings were observed in both the transformers and BP network, whether trained on both synthetic languages A & B or independently. The paper also proves, with the proposed theory, that there is a better representation to present the shared semantic content by subtracting the predictable component from preceding layers, which is called layerwise novelty. Also, separates coordinate alignment from geometric similarity affected by the code-switching rate and training size via latent cross-probing.

Critical strengths
  • The layerwise novelty part is a novelty and a significant contribution in terms of LLM theory and interpretability. This highlights the applicability of the theory proposed by the paper, and the paper itself is able to prove it empirically.
  • Able to distinguish the geometric similarity in multilingual representations and coordinate alignment for cross-lingual transfer, which is I believe critical in terms of understanding LLM dynamics around this area.
Critical weaknesses
  • Incremental extension from the monolingual RHM and BP kinda limited the contributions of the work
  • The way the experiment is conducted does not explain further on why the training dynamics on transformers, mainly on gradient descent, map monotonically to layer depth. What I mean by it is the multilingual RHM data and the usage of BP network kinda induce the middle-layer similarity peak directly into the network construction. Knowing how BP message passing work, it is kinda intended and expected to behave this way. Maybe I should say this is more towards algorithmic proxy rather than theory of how transformer learns multilingual data?
  • I think modelling cross-lingual divergence solely as disjoint symbol substitutions on a fixed-depth oversimplified actual linguistic structure.
Questions for authors
  • If the languages differ in a more complex way, for example like order (one is right-to-left, other is left-to-right), or any other complex differences, do most findings still hold?
  • can the authors clarify why the standard transformer is expected to follow the same approach as BP?
2 comments
AuthorsOpenReview ↗

Response to Reviewer QHDj

We thank the reviewer for their thoughtful review.

Contribution beyond RHM: The technical contributions of our paper going past previous works include studying next-token prediction on the Random Hierarchy Model (RHM), and constructing theoretical networks that solve this task using encoded belief propagation (BP) messages. Most importantly, our paper introduces a framework to model different languages/modalities, based on the notion that they share the high level rules of their generative hierarchy. This key addition allowed us to capture many phenomena observed but unexplained in LLMs, and also make new predictions which we verified.

Training dynamics vs algorithmic proxy: A full-fledged theory of gradient descent in deep architectures (let alone transformers) is a long-standing question in the field, which has not been resolved. However, the RHM is one of the rare cases where a rigorous theory of dynamics approximating GD was made in MLPs, see Ren et al (2026). It corresponds to the supervised setting where the root is predicted, and where the networks learn a representation similar to the encoding part of BP.

This is consistent with our results, in particular Fig.10 that provides a direct check of layerwise agreement between the theoretical BP network and the trained transformer. We show that the accuracy of linearly predicting layer-ll BP representation hBP(l)h^{(l)}_{BP} from layer-l′l' transformer representation htransformer(l′)h^{(l')}_{transformer} follows the same trend as predicting hBP(l)h^{(l)}_{BP} from hBP(l′)h^{(l')}_{BP}.

We will discuss the existing connections between our theoretical network and GD, and cite Ren et al.

Complex linguistic structures: Any tractable theory of natural languages must obviously make simplifying assumptions. Our goal is to use the simplest setup that captures various observations in LLMs. The comparison with natural languages indeed support that our conclusions are robust to the specific choices we made. Robustness is already tested with respect to our choice of fixed topology trees: in Fig.17 of the paper, we show that our conclusions also hold for trees with variable topology. In response to the specific comment of the referee on shared v.s. non-shared vocabulary, we show in https://anonymous.4open.science/r/rebuttal_shared_vs_disjoint_vocabulary-04FD/ that having partially shared, instead of disjoint, vocabulary does not qualitatively change our conclusions.

Languages with Left-to-right vs right-to-left order: The transformer's next-token prediction objective follows the order in which the text is read, regardless of whether the script is written left-to-right or right-to-left. We thus expect our findings not to depend on script direction.

References:

Yunwei Ren, Yatin Dandi, Florent Krzakala, Jason D. Lee, Provable Learning of Random Hierarchy Models and Hierarchical Shallow-to-Deep Chaining. Colt 2026.

Thank you for the very insightful response. I think now my concerns has been address, I will adjust my review accordingly.

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

This paper proposes a theoretical framework to explain cross-lingual representation similarity in language models. The authors introduce a Multilingual Random Hierarchy Model (MRHM), where languages share high-level hierarchical structures but differ in their surface-level production rules. By modeling Bayes-optimal next-token prediction through belief propagation (BP), the framework explains why cross-lingual similarity peaks in intermediate layers and increases with language proximity and training data exposure. The paper further distinguishes geometric similarity from coordinate alignment, showing how code-switching promotes the latter. Finally, it introduces laye-rwise novelty, which improves cross-lingual representation similarity by removing information linearly predictable from preceding layers, with supporting experiments on synthetic data and pretrained LLMs.

Critical strengths
  1. The paper provides an elegant and interpretable framework that explains several empirical observations about cross-lingual representations.
  2. The proposed layer-wise novelty method is a particularly interesting contribution. Its effectiveness is demonstrated on both synthetic data and pretrained LLMs.
  3. The paper clearly distinguishes geometric similarity from coordinate alignment and provides insights into how code-switching facilitates cross-lingual alignment.
Critical weaknesses
  1. The framework assumes that different languages share high-level generative structures while retaining language-specific lower-level structures. While the resulting agreement with trained transformers is interesting, it remains unclear to what extent the theory explains the emergence of these representations rather than reproducing a pattern largely determined by its assumptions.

  2. While the controlled setting of training a transformer on CFG-generated data is interesting and clean, a similar U-sharp observation doesn't mean that real LLMs necessarily work in the same way. I would expect the authors to provide a bit more evidence to align the behavior. Coordinate alignment requires explicit cross-lingual signals, which seems to be aslo a strong assumption. In reality, alignment could probably appear based on other signals, such as numbers.

  3. More related references can be added to enhance the background. I list two relevant ones, but definitely not limited to them. Tang et al. Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models. ACL 2024. This paper identifies language-specific neurons across layers. A similar U-shaped distribution is observed. Zhong et al. Language Lives in Sparse Dimensions: Toward Interpretable and Efficient Multilingual Control for Large Language Models. EACL 2026. This paper also identifies language-agnostic and language-specific dimensional spaces from the representations. They perform subtraction between the middle and final layers to control the generated languages.

Questions for authors

see weaknesses

3 comments
AuthorsOpenReview ↗

Response to reviewer tNmi

We thank the reviewer for their thoughtful review.

Assumptions and scope of our theory: "All models are wrong, but some are useful". The usefulness of a scientific theory is determined by the sparsity and naturalness of its assumptions, how it explains known phenomena, and how it predicts new ones.

As noted by the reviewer, our theory assumes that different languages and modalities share high-level structure in the generative hierarchy. This single axiom allows us to capture many known phenomena in LLMs, including the coexistence of shared and language-specific structure, the layerwise organization of geometric similarity, increased similarity with greater language proximity and higher model quality, similarity trends for low-resource languages, and coordinate alignment and transfer across languages. It also allows us to make a new prediction that layerwise novelty improves the geometric similarity of representations. We verify this prediction in LLMs and natural languages. This extent of empirical evidence supports that our theory will be of the useful kind.

This also addresses the reviewer’s comment that “a similar U-shaped observation doesn’t mean that real LLMs necessarily work in the same way.” We again emphasize that the U-shaped layerwise similarity trend is one of several points of agreement between our theory and real LLMs. The evidence includes all the observations listed above, as well as our empirical verification in LLMs of the new prediction that layerwise novelty improves the geometric similarity of representations.

Source of coordinate alignment: We focus on mixed-language sentences as one mechanism for coordinate alignment. The reviewer is right that other mechanisms may also contribute to alignment. Among others, this includes shared vocabulary, such as numbers, that may act as anchors. In our setup, this mechanism could be implemented by allowing the languages to share some rules at the bottom level of the hierarchy. Comparing extensively all possible mechanisms for language alignment, and studying how they interact, is an interesting question that clearly requires a work of its own.

Additional references: We thank the reviewer for bringing these highly relevant works to our attention. We have cited both works in the revised version of the paper. The work by Tang et al (2024) is especially very interesting, since they find language-specific neurons in early and late layers, in line with our theory. It is likely that such language-specific neurons also exist in our setup -- this is an interesting question for future exploration.

As the discussion window is coming to a close, we kindly ask whether we have addressed all of the reviewer's questions and concerns, and whether we can provide any further clarifications.

Thank you for your response. A part of my concerns have been addressed, and I believe my current positive score appropriately reflects my assessment.

Reviewer_zg13OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

The authors contribution is to show that modeling multilingual LLM as processing languages with a hierarchical structure predicts several results from previous work e.g. LLM representations similarity across languages peak in middle layers. They first validate this hypothesis in a simple synthetic case by constructing and analyzing the representation of a network implementing the known optimal algorithms, and validate that transformers trained on this data show similar trends. They then move to real world LLM and show that for multiple language the same predictions hold. This is interesting because it unifies a lot of previously observed property of multilingual LLMs into a single assumption on which future work can build on.

Critical strengths
  • Strength 1: Proposes a model of how LLMs learn to represent different languages that predicts a lot of observation seen in the literature
  • Strength 2: Nice progression from theoretical network to small transformer to LLM, while the writing is not perfect and some important bits are in the appendix, I found the paper pleasant to read and very well structured with 0% if sloppy / hard to understand writing.
Critical weaknesses
  • Weakness 1: While the paper's conclusion is that layerwise novelty is a Novel and promising way to study representations, the current main body doesn't support that imo. layerwise novelty is basically just an affine prediction from layer l-1 to l: this is only explained in appendix I, I think the main text should contain at least an intuitive description of how it's implemented. Thus it is not entirely intuitive for me why this would be superior than just looking at the residual contributions of each layer. Figure 18 (also in appendix) show that just the residual contribution is as good as novelty for their toy models. Figure 19 expand this comparaison to real LLMs but is just too hard to analyze and conclude anything, (although novelty seems slighly better maybe?).* A lot of the claim about novelty being better use a log scale of II and the authors do not explain why from figure 1 to 4 we have linear scale and from 5 on log scale. Figure 6 doesn't have confidence intervals as far as I can tell. I am not familiar with II, but it seems reasonable to use log scale as iiuc 0.1 to 0.01 means the average rank goes from top 5% to top 0.5%, but yeah it's unclear whether such a difference would be significant and justify using layerwise novelty over simpler metrics.
  • Weakness 2: the framing that this is an advance on the platonic representation hypothesis is a bit weird to me: almost all experiment compare representations of the same model across languages. The "Separately trained models also exhibit geometric similarity." is predicted by the theory but only validated in the toy setup, not on real LLM
  • Also the choice of log scale for mKNN seems inappropriate here, as it compresses the 0.1, 1 range which is significant while 0.1 to 0.01 is terrible in both cases
Questions for authors
  1. Why do you switch to log y axis for II for some of the plots? Why should I find a 0.01 vs 0.1 II score significant? I asked LLMs about previous literature but didn't get any prior work showing that this would correlate with a qualitative difference in representation? Does this question actually make sense or am I missing something?
  2. Are fig 6 CI negligible? could you make sure to include them

typos and remarks

As ICLR don't have a section for typos / other slightly confusing stuff, I include some notes here in case they are useful to authors. No need to respond to this in your rebuttal as those are not load bearing (smiled while typing that) for my score:

  1. Fig 8 seems to be wrong it says X^(1) even at the root which should be X^(4) (on the right in blue)
  2. L390 "We call this residual layerwise novelty and predict that it exposes shared computation more clearly than the raw hidden states.". This was not intuitive to me but L396 made it make sense "In later layers, novelty shows higher II because it removes the high-level information carried by the skip connection"
  3. Fig 5: II results on empirical network is so clean! you have as many low II layers as L_shared, somehow this is not the case in the BP network. I'm guessing this is just an artifact of the way BP is built?
  4. Fig 6: could be cool to relate layers here to their role according past work like Wendler et al. [1], Dumas et al. [2] and Tang et al. [3]. E.g. does II drop in layers where logit lens is english [1]? Does II drop in layers in which patching starts to transfer concepts [2]? Does II rise in the early and late layers, where language-specific neurons are concentrated [3]?
  5. Who is Fig 1.b for? I was not familiar with RHMs but familiar with grammars and this diagram contributed mostly to confuse me, while your description in the caption was pretty clear. On my first read even the figure in Appendix A didn't really make it intuitive how text is generated from an RHM. I ended asking claude for a video explainer and then everything was clearer, feel free to steal / inspire yourself from it https://anonymous.4open.science/r/rhm-explainer-anon-CC79/video/rhm_explainer_720p.mp4

[1] Chris Wendler, Veniamin Veselovsky, Giovanni Monea, Robert West. Do Llamas Work in English? On the Latent Language of Multilingual Transformers. ACL 2024. arXiv:2402.10588 [2] Clément Dumas, Chris Wendler, Veniamin Veselovsky, Giovanni Monea, Robert West. Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers. ACL 2025. arXiv:2411.08745 [3] Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Wayne Xin Zhao, Furu Wei, Ji-Rong Wen. Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models. ACL 2024. arXiv:2402.16438

3 comments
AuthorsOpenReview ↗

Response to Reviewer zg13

We thank the reviewer for the thoughtful and detailed review.

Comparing layerwise novelty with residual contributions: Compared to naive residual contributions, layerwise novelty is a more robust method to isolate high-level signal in representations, as it can handle rotated/dilated representations. While the two methods are equally good in our synthetic setup (Fig.18), this distinction is important in practical setups: in natural language layerwise novelty performs much better than residual contributions (Fig.19). In Fig.19, layrwise novelty (dashed lines) improves similarity while residual contributions (dash-dot lines) makes it worse.

We agree that layerwise novelty can be described more clearly in the main text. In the revised version of the paper, we do so at the beginning of Section 6.

Log scale of information imbalance: As correctly pointed out by the reviewer, information imbalance (II) going from 0.1 to 0.01 means that for 1000 evaluation pairs, the average rank goes from 50 to 5, which is a significant improvement. Intuitively, this means that the two representations agree on which sentences are close at a much finer scale. Therefore, the improvements with novelty are indeed significant, and a logarithmic scale best highlights that.

Figures 1-4 remain qualitatively unchanged in both linear and log scales. For consistency, we have now changed all the figures to always show II on a logarithmic scale.

Linear scale of mKNN: We agree with the reviewer that linear scale is more appropriate for mKNN. We have updated Figures 18 and 19.

Geometric similarity within the model vs across models: As pointed out by the reviewer, our theory indeed predicts that the geometric similarity trends are the same both within a model and across separately trained models. In our synthetic setup we empirically verify this in trained transformers in Fig.2. Thus, our setup indeed captures the platonic representation hypothesis. For LLMs, geometric similarity within and across models is already observed in the literature, for example in Fig.4 of Acevedo et al (2026). For completeness, we have extended Fig.6 to include cross-model similarity results, which includes raw representations and layerwise novelty.

CIs in Fig.6: In the revised version of our paper, we have updated Fig.6. We now average over multiple data samplings for II measurements, and include the error bands.

References:

Santiago Acevedo, Andrea Mascaretti, Riccardo Rende, Mateo Mahaut, Marco Baroni, and Alessandro Laio. "A quantitative analysis of semantic information in deep representations of text and images". TMLR 2026.

Reviewer_zg13OpenReview ↗

Raising my score to 4

The authors have addressed all my concerns, in particular the cross-model comparaison is very clean.

I have revisited Fig 19 and agree that layerwise novelty improves clearly on residual contributions, I'm unsure if the authors changed the figure or if I struggled more at the time of writing the review.

I am therefore raising my score from 3 to 4 and presentation and soundness from 2 to 3.

Thank you!

We are glad to have addressed all of reviewer's concerns.

We did not change Fig.19 beyond the yscale of mKNN. We will modify it with different colors for different curves to make the dense plots clearer.

We again thank the reviewer for all the helpful comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.