acceptodds
Under review as a conference paper at ICLR 2027

The Shared Geometry of Language Models

Abstract

Language models with different sizes and architectures develop strikingly similar representation geometry, with specific observed arrangements such as cycles and helices often interpreted as traces of computation. We propose the simpler explanation that much of this shared geometry reflects the geometry at model input and output. We show that the shape of input embeddings extends into middle layers while representations converge to the geometry of their predictions well before the last layer. To understand what drives the geometry of these input and output weights, we train models without hidden layers or with only one attention layer and show that they recover a large fraction of the geometry shared across models. This suggests that low-order corpus statistics are a major source of the shared input and output geometry. We support this claim with tokenizer interventions that produce predictable geometric changes in both shallow and deeper models. Finally, we trace six previously reported structures in LLMs to the shape of the input or the shape of the answer. This suggests input and output geometry as an alternative explanation to rule out before attributing representation geometry to task-specific computation.

Primary area:
interpretability and explainable AI
Keywords:
Representation Geometry, Feature Geometry, Feature Manifolds, Mechanistic Interpretability

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_X5hXOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2(+1)
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper proposes that representation geometry seen across models is driven by input and output geometry. The paper demonstrates that model embeddings and unembeddings share structure across models based on dot-product kernels. They also show that a 1-layer model can recover much of this structure suggesting that embedding and unembedding geometry is driven by simple statistics. They then attempt to attribute intermediate representation geometry to either the input or output or both.

Critical strengths
  • The use of kernel geometry and global metrics like CKA in addition to local metrics like mean neighborhood overlap help provide support for the findings about the shared geometry and ensure that the results are not simply due to high-variance directions.

  • The comparison with shallow models supports the idea that the embedding and unembedding geometry are primarily low-order statistics.

Critical weaknesses
  • The main concern is the fact that it is unclear whether the key claim that representation geometry can be mapped back to the embedding or unembedding advances our understanding of model representations and is novel. The fact that representations towards the beginning have similarities with the initial embedding follows from the residual connections and the fact that the model must transform this embedding by adding on new components at each layer. Similarly, the fact that representations towards the end have similarities with the final unembedding follows from the fact that the model aims to eventually decode its representation into a token. As a result, it is unclear whether explaining intermediate representations in this manner provides new insights, especially when Figure 2 demonstrates that agreement with representations at either end is quite low in the middle of the network, and such relationships between low-order statistics and representations have been explored in works such as [1]. Additionally, the analysis is primarily focused on settings where the concepts/structures are primarily based on single words/tokens and may have limited applications beyond this scope. Furthermore, analysis such as the one in Section 5.1 seem to simply be a verification that the output representation more strongly appears after layer 18 which was shown to contain key addition neurons in previous work, and it does not seem to add any additional explanation for how these representations transformed beyond the fact that the model eventually needs to decode the answer.

[1] Symmetry in language statistics shapes the geometry of model representations, Dhruva Karkada, Daniel J Korchinski, Andres Nava, Matthieu Wyart, Yasaman Bahri

Questions for authors
  • Could you provide clarification on how this work adds to the existing literature such as works using LogitLens or works that analyze the relationship between low-order statistics and representations?
5 comments
AuthorsOpenReview ↗

Author response

We thank the reviewer for their time and feedback. Below we make the case that our work is novel in providing a simple explanation of observed representation geometry that has predictive power over the geometry of early and late layers of real LLMs and is validated with quantitative metrics beyond specific examples like months or integers. We also argue against the notion that these results 'follow' from residual connections and note that connection to embedding and unembedding geometry was not considered as a hypothesis despite extensive work on this topic over the past 2 years.

We position our work with respect to two existing literatures:

1) Works that identify geometry/manifolds in LLM representations: Engels et al. (2025), Kantamneni & Tegmark (2025), Wurgaft et al. (2026), Feucht et al. (2026), Park et al. (2025a), Modell et al. (2025).

2) Works that link these structures to low-order statistics: Karkada et al. (2026), Prieto et al. (2026).

Relation to (1). None of the works in (1) consider the embedding or unembedding geometry as an explanation for the structures they find, even those that focus on early and late layers. They focus instead on what the geometry might do computationally, or on what it implies for the linear representation hypothesis and for steering. Wurgaft et al. explicitly leave the origin of the representation–behaviour correspondence outside their scope. Understanding where such geometry comes from is listed among the main open problems of mechanistic interpretability (Sharkey et al., 2025).

Two of our novel empirical findings from this work make the connection to embedding and unembedding geometry worth testing:

a) Most of the previously studied cases of representation geometry/manifolds in the hidden representations of LLMs are already present in the embedding and unembedding.

b) The same structures are also present in the embedding and unembedding of 1-layer models.

Why this does not simply follow from residual connections and LogitLens.

  • Output side. Linking geometric arrangement in late layers to the shape of the answer requires leaning on the empirical LogitLens results and the nature of residual connections, but also knowing that the structure is present in the unembedding and understanding how unembedding structure determines the pre-image of the target logits (Figure 4). We see our result as the geometric counterpart of the LogitLens when combined with structured unembeddings.
  • Input side. No analogue of the LogitLens predicts that embedding geometry should be preserved through early layers rather than quickly overwritten. Residual connections make it plausible that some input information persists but do not determine which relational structure persists, how strongly, or for how long, which is what we measure (Appendix D). For example, it was plausible given our previous results, but not necessary, that the structures captured by SAEs in layer 7 by Engels et al. (2025) would also be present in the embeddings and in every layer in between (Figure 11).

With respect to the works in (2), we believe our work makes several important contributions. While Karkada et al. (2026) and Prieto et al. (2026) establish a connection between low-order statistics and representations, their main objects of study are classical word embeddings like word2vec (for Karkada et al., 2026) and shallow autoencoders (for Prieto et al., 2026). Their connection to language models is limited to noting that a few of the same structures in representations of specific concepts like months or years look qualitatively similar between LLMs and word embeddings. We build on this by showing that even the embeddings and unembeddings of LLMs with billions of parameters are driven by similar statistics and measure this agreement using quantitative metrics like CKA across the full vocabulary, beyond specific word groups like months or days. While works in (2) show the appearance of similar structures to (1) in simpler settings, we replicate the results from (1) and show that they can be mapped to the shape of the input or the shape of the answer in real LLMs.

AuthorsOpenReview ↗

Author response (part 2)

Middle layers. While low-order statistics do not explain all representation geometry, we show that they explain a surprisingly large fraction of the geometry shared by different language models, without trying to explain the geometry that is not even shared by different LLMs. We have now made this clearer and added a new line to Figure 2 showing the agreement between the hidden representations of two different LLMs as a reference. Even taking this baseline into account, it is true that a larger fraction of geometry is unexplained in the middle layers. This unexplained geometry in middle layers that is shared across models is an interesting place to look for structure that appears with scale and reflects more sophisticated computation. We think this is another way in which our work advances our understanding of model representations by discarding structure that has simple explanations and allowing interpretability researchers to focus on structure that is not explained by the shape of the input or the shape of the answer.

Section 5.1. The goal of Section 5.1 is not to relocate the addition computation, which Feucht et al. (2026) identify at the layer-18 MLP. It is to explain an observation their account leaves open. Feucht et al. find that the period-12 month circle is absent where the computation happens: “we cannot reliably probe for circular structure at the final token position until layers 22-25 … This suggests that the causal mechanism at layer 18 does not rely on task-specific circular geometry.” If the circle is not used by the computation, why does it appear at all? We answer this by showing that the late circle is the shape of the answer: since the residual stream must produce month logits through the unembedding, it inherits the cyclic arrangement of the month unembedding rows. Consistent with this, answer geometry overtakes offset geometry around layer 19, and prompts are organised by answer month at layers 22 and 29 (Figure 3), matching the layers where Feucht et al. recover the circle. The location of the circle is therefore predicted by where the answer forms, not by where the computation happens.

We hope our response has alleviated the reviewer's concern with novelty and would allow them to increase their score. We remain available for the remainder of the rebuttal period to address any further questions or concerns.

References

  • Engels, J., Michaud, E. J., Liao, I., Gurnee, W., & Tegmark, M. (2025). Not All Language Model Features Are One-Dimensionally Linear. International Conference on Learning Representations (ICLR). https://openreview.net/forum?id=d63a4AM4hb
  • Feucht, S., Haklay, T., Bhalla, U., Wurgaft, D., Rager, C., Sarfati, R., Merullo, J., McGrath, T., Lewis, O., Lubana, E. S., Fel, T., & Geiger, A. (2026). Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Concepts. arXiv preprint arXiv:2605.01148. https://arxiv.org/abs/2605.01148
  • Kantamneni, S., & Tegmark, M. (2025). Language Models Use Trigonometry to Do Addition. arXiv preprint arXiv:2502.00873. https://arxiv.org/abs/2502.00873
  • Karkada, D., Korchinski, D. J., Nava, A., Wyart, M., & Bahri, Y. (2026). Symmetries in Language Statistics Shape the Geometry of Model Representations. International Conference on Machine Learning (ICML). https://arxiv.org/abs/2602.15029
  • Modell, A., Rubin-Delanchy, P., & Whiteley, N. (2025). The Origins of Representation Manifolds in Large Language Models. arXiv preprint arXiv:2505.18235. https://arxiv.org/abs/2505.18235
  • Park, C. F., Lee, A., Lubana, E. S., Yang, Y., Okawa, M., Nishi, K., Wattenberg, M., & Tanaka, H. (2025a). ICLR: In-Context Learning of Representations. International Conference on Learning Representations (ICLR). https://arxiv.org/abs/2501.00070
  • Prieto, L., Stevinson, E., Barsbey, M., Birdal, T., & Mediano, P. A. M. (2026). From Data Statistics to Feature Geometry: How Correlations Shape Superposition. International Conference on Learning Representations (ICLR). https://arxiv.org/abs/2603.09972
  • Sharkey, L., Chughtai, B., Batson, J., Lindsey, J., Wu, J., Bushnaq, L., Goldowsky-Dill, N., Heimersheim, S., Ortega, A., Bloom, J., et al. (2025). Open Problems in Mechanistic Interpretability. Transactions on Machine Learning Research (TMLR). https://arxiv.org/abs/2501.16496
  • Wurgaft, D., Rager, C., Kowal, M., Shyam, V., Feucht, S., Bhalla, U., Haklay, T., Bigelow, E., Sarfati, R., McGrath, T., Lewis, O., Merullo, J., Goodman, N. D., Fel, T., Geiger, A., & Lubana, E. S. (2026). Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior. arXiv preprint arXiv:2605.05115. https://arxiv.org/abs/2605.05115

Thank you for the in-depth response and the clarifications! I appreciate the clarifications around the related works and section 5.1.

While I also appreciate the shared baseline to demonstrate the shared structure ceiling, I believe this remains a limitation of the input or output shape explanation itself.

The main concern that remains is that it is unclear whether the proposed idea goes beyond identifying correlations between the embedding/unembedding structure and structure from previous findings. While this provides a quantitative measure for identifying how similar the input/output structure is to intermediate embeddings, it remains unclear whether interventions on the shared parts of the representations result in the input or output-level representation shifting similarly and more generally, how it should be used as a baseline explanation. Furthermore, in Figure 3, it seems that the circular structure actually has more noise and explains less of the variance at layer 29 than at layer 25, which makes it uncertain how strongly the answer structure corresponds to the structure being studied such as the circular months.

I do appreciate all of the additional results and explanations and do believe the informative and soundness have been strengthened, and will raise my score from 1 to 2.

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

The paper calls into question recent results from mechanistic interpretability, which finds (1) representational manifolds in the deep representations of LLMs; and (2) connects this manifold geometry to the computation being performed. The authors do this by showing that the manifold geometry in the deep layers of LLMs can be largely explained by shallow surface statistics learnable by a 0- or 1-layer language model. Thus, representational geometry in the deep layers of LLMs may not reflect computation, but rather, simply input / output embedding geometries.

Critical strengths

Kudos to the authors for writing an exceptionally clear, and easy to follow, paper. I genuinely enjoyed reading it.

Another clear strength of the paper is the question it tries to tackle, i.e., whether "shared neural representational geometry" of concepts across models can be simply attributed to geometry of the embedding / unembedding, rather than some learned structure. This question is quite ambitious, and potentially quite impactful to the community as it engages with several known results (e.g., weekdays, numbers, etc).

Critical weaknesses

I have concerns with the primary claim (copy-pasted below):

Our account traces disparate observations of representation geometry to a common origin, identifying previously unexplained structure in prior work as either the shape of the input or the shape of the answer. It thus provides an alternative explanation to test before attributing observed structure to task-specific computation.

My understanding is the following. Throughout the paper, the main argument is that previous attempts to attribute representational geometry to "task-relevant" computation can be explained due to low-order statistics in the input. Thus, this would diminish previous claims.

Here, I disagree with the authors. Learning low-order statistics is part of the "task" to predict natural language (which drives learning to represent concepts "informatively", e.g., along the curved number manifold). To me, the fact that single-layer or no-layer LLMs recover much of the representational geometry would instead indicate that the domains (numbers, years, months) are too simple, and can be learned with a small model capacity.

In order to build a more convincing argument that it's truly surface-level statistics driving representational geometry, I would like to see the results replicated on a random-feature control (e.g., by taking RFF embeddings on the input, or a random linear map to R^768), as well as on a much more difficult task.

If the authors can resolve my concern (and incorporate the discussion on Related Work), I will raise to a Weak Accept.

Questions for authors

Missing citations:

Cheng et al., (ICLR 2025) puts forth a very similar layerwise analysis on similarity to the first and last layers. They and Wolfram & Schein, (COLM 2025) find that the middle layers are most similar between LLMs, which should be discussed in related work. The fact that middle layers tend to be most similar between models is evidence contra your conclusion that intra-model similarity can be explained by input/output embedding geometry. What, in your view, accounts for this discrepancy?

2 comments
AuthorsOpenReview ↗

Author response

We thank the reviewer for their feedback as well as for complimenting the clarity and ambition of the paper. Below we address the reviewer's main concerns:

What we mean by 'task relevant computation': We agree that capturing low-order statistics is a key part of the 'task' of language modelling. The question we aim to tackle is whether these geometric structures appear in order to enable sophisticated computation involved in specific sub-tasks like arithmetic or whether they are due to much simpler demands from modelling low-order statistics. Our wording “task-relevant computation” blurred this distinction, so we have revised the text to say “computation involved in specific sub-tasks” across the paper.

Task difficulty: The reviewer points to an important notion that we now make clearer. One of the main points we want to convey is that embedding geometry can be simple even when the computation is hard. 1-layer models encode integers similarly to LLMs despite not being able to perform almost any numerical tasks. We now include more detailed results (Appendix B.4) showing that the 1-layer model can only perform the simplest numerical tasks like matching an integer to the right ordinal suffix ('1st, 2nd, 3' ->'rd') with accuracies below 10% on other tasks like addition or subtraction which Llama solves with 100% accuracy. 1-layer models have not learned most of the tasks that involve numbers or months as variables but they have learned enough of the simple regularities related to these variables to converge to similar arrangements in the input and output mappings.

Regarding the random feature control, we understand the reviewer's request to be regarding a check that the geometric agreement is not an artifact of our chosen metrics by comparing to a random embedding baseline. We now show in Figure 1 that a random embedding baseline achieves agreement near 0 with both of our metrics in all the settings evaluated. If this was not the concern the reviewer intended to express, we are happy to address this issue further if they can provide additional clarification.

Missing citations: We thank the reviewer for bringing these relevant works to our attention, a discussion of their results has now been included in the related work. Our claim is not that all shared LLM geometry is explained by input/output geometry but rather that it is an important driver of representation geometry that must be ruled out before exploring other possible origins for these structures. We have updated Figure 2 to show that cross-model agreement is stable across layers when using our metrics, in line with Wolfram & Schein (2025)(§4, Fig. 5) but also discuss in the related work that agreement seems to increase in middle layers when using the Information Imbalance metric from Cheng et al. (2025).

Independently of the choice of metric, it is clear that some of the shared geometry in middle layers is not captured by either the shape of the input or the shape of the answer. While we show in Section 5 that most previously studied representation manifolds can be explained by endpoint geometry, the unexplained fraction of shared geometry in middle layers (corresponding to layers with high intrinsic dimension in Cheng et al. 2025) suggests the existence of shared structure that appears only in larger language models. It is precisely because we think hidden layers can contain rich structure that is informative about the inner workings of LLMs that we emphasize the need to be aware of endpoint geometry as an alternative explanation to check before looking for more sophisticated functional roles for this kind of geometric arrangement.

We hope our response has alleviated the reviewer's concerns and remain available during the rest of the rebuttal period for any further questions or concerns.

Thanks for the clarifications! Your experiments are exactly what I meant with the random-feature baseline. I'll raise my score to a 3.

Reviewer_4v1jOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates the origins of shared representation geometry across language models, proposing that much of this structure arises from input and output interfaces shaped by low-order corpus statistics rather than complex computations in deeper layers. The authors show that zero- and one-layer models recover much of the shared embedding and unembedding geometry and examine how these endpoint structures relate to hidden representations across layers. Through tokenizer interventions and analyses of previously reported geometric patterns, they provide evidence that certain structures can originate from local prediction statistics and persist into deeper representations.

Critical strengths

The paper addresses an important question about the origins of shared representation geometry across language models and offers a simple, well-motivated intuition. If structures attributed to deep computation already appear in the embedding and unembedding of zero- and one-layer models, complex computation may not be necessary to explain their origins. Tracking how this endpoint geometry persists into hidden layers provides a concrete, testable alternative explanation and a useful null hypothesis for interpretability claims.

Critical weaknesses

1. Paper organization and clarity of presentation

The paper's organization and writing need substantial improvement, as they make it unnecessarily difficult to follow the main arguments and understand the experimental results.

In particular, the introduction does not provide a sufficiently clear and informative overview of the proposed approach and experimental methodology. Although it introduces the central hypothesis, the reader must wait until Sections 2 and 3 to understand the experimental setup and how the authors investigate their claims. The paper should also define its terminology more explicitly from the outset. For example, it should clarify that the shallow models refer to the zero-layer and single-attention-layer baselines.

The methodology is explained more clearly in later sections, and its connections to previous work are generally well established. However, several presentation issues remain:

  • Figure 1: The term shallow models should be defined before or alongside its first use. Furthermore, Figure 1A is not sufficiently self-contained; the figure and caption should more clearly explain the reported prediction-loss comparison.
  • Appendices: The main text does not provide a clear overview of the contents and purposes of the different appendices, apart from the more explicit discussion of Appendix D.
  • Figure explanations and consistency: Several figures are insufficiently explained, making their results difficult to interpret without consulting additional sections. There are also inconsistencies between the figures and the main text. For example, line 421 refers to "Figure 5D–E," whereas Figure 5 contains only panels A and B.

Overall, the presentation should be revised to make the paper's contributions, methodology, and experimental findings easier to follow.

2. Insufficient evidence that shared unembedding geometry explains late-layer alignment

The paper argues that shared late-layer geometry partly arises from similar unembedding matrices. However, the shape of the answer, defined by pulling each model's own final logits back through its unembedding, recovers its final hidden state (Appendix A.3). Thus, the corresponding curve in Figure 2 primarily shows that late-layer representations increasingly resemble the final state, rather than explaining the origin of this geometry.

The more informative test uses Mistral-Nemo's logits pulled back through the one-layer model's unembedding. Yet randomly permuting the unembedding's token rows, breaking token-to-readout correspondence, only modestly reduces agreement (CKA: 0.571 to 0.517; neighbors@10: 0.298 to 0.282, Appendix C.5). This suggests that much of the alignment on natural text comes from prediction structure rather than the correctly aligned unembedding interface. Figure 4B provides clearer evidence of an interface effect in a constructed weekday task, but the broader claim that shared unembedding geometry explains late-layer alignment is not sufficiently supported.

3. Endpoint geometry does not fully explain shared representation geometry

The paper claims that "much of this shared geometry" reflects input and output interfaces and that these endpoints provide "a simple account of the origin of representation geometry." However, Figure 2 only measures agreement with each endpoint separately. In middle layers, both references show relatively low agreement with hidden-state geometry (approximately 0.3 CKA and 0.2 neighbor overlap).

Although residual connections and the appendix experiments support the persistence of specific structures, such as calendar circles, this does not establish how much of the full hidden-state geometry the two endpoints jointly explain. As Appendix C.3 acknowledges, these measurements are not a decomposition of cross-model CKA or explained activation variance. The evidence therefore supports endpoint contributions to representation geometry, but not the broader explanatory claims.

4. Limited evaluation of tokenizer dependence

The shallow-model recovery experiments (Section 3) and tokenizer interventions (Section 6) rely on the GPT-2 tokenizer, while the main cross-LLM comparison is restricted to 9,635 tokens that are single tokens across all evaluated tokenizers. This is particularly concerning for integer geometry, where the authors compare against Pythia because other models tokenize integers differently.

The authors should repeat the shallow-model recovery and numeral intervention using at least one additional tokenizer, ideally with single-digit tokenization and a different vocabulary size, and report the same period-10 Fourier-energy and consensus-agreement metrics.

Questions for authors
2 comments
AuthorsOpenReview ↗

Author response

We thank the reviewer for their time and useful feedback. Below we address the reviewer's main concerns.

  1. We have included the reviewer's suggestions and made additions to the paper to improve clarity. Concretely, the introduction now gives more methodological detail (L51-L53, L57-58) and avoids the use of terms like 'shallow models' before they are defined. We also fixed inconsistencies like Figure5 D-E, extended the captions to make them more self-contained and included additional references to the Appendix from the main text to give the reader a better idea of what it contains.

  2. We agree with the reviewer that prediction structure is an important component of the shape of the answer. As discussed in Section 5.2, the shape of the answer is jointly determined by the shape of predictions and the interface through which predictions are being made. We believe the unembedding interface becomes more important for the shape of the answer when target predictions carry little structure, for example in well defined taks with a single correct answer such as our weekday example in Figure 4B. We validate this hypothesis by showing that learned unembeddings (coming from deep or shallow models) predict late layer geometry significantly better than a random permutation in a new person-birthplace task. At the last layers, both the models’ own unembedding rows and those of the 1-layer model reach CKA 0.63 with birthplace geometry, against a 95% permutation null below 0.06 (Appendix C.6). These results are now included in Appendix C.6 and referenced in the main text.

  3. Our claim refers to the geometry that is shared across LLMs, however this was unclear in our submission as that reference was missing from Figure 2. We have added a new line to Figure 2 showing how much geometry is shared between hidden layers of different LLMs as a baseline. We also measure the agreement with the combined shape of the input and answer as suggested by the reviewer (Appendix C5). At half depth, the joint reference reaches held-out CKA 0.50 (neighbors@10 0.24), compared with 0.35 for the input and 0.42 for the answer alone. Pairwise agreement between different LLMs at the same depth is 0.69 (neighbors@10 0.36). Near the endpoints, the references match the cross-model level: 0.77 for the input against 0.76 pairwise after the first layer, and 0.60 for Nemo’s logits pulled back through the 1-layer unembedding against 0.74 pairwise at the last. This makes it clear that the shape of the input and the shape of the answer do explain a large amount of shared geometry. They do not explain the geometry that is not shared across LLMs , but more importantly, we see that some geometry that is shared across LLMs remains unexplained in middle layers, we believe this is a promising place to look for representation geometry that appears with scale due to more sophisticated computation rather than low-order statistics.

  4. We thank the reviewer for pointing to this oversight, the only model that needs to be excluded from the integer plot in section 3 is Qwen since Llama OLMo and Pythia all have different tokens for integers 1-99, we have added these 2 additional models in Figure 1 shown the same integer spectrum appears in the consensus of these 3 models. Regarding the use of 9,635 tokens, this only applies to the nearest neighbor results in Figure 1, we provide frequency stratification results for CKA across the full shared vocabulary. We have now updated the paper to do the same for nearest neighbors in Appendix B2. With these changes we show that our results are not an artifact of comparing only to Pythia or evaluating only on 9,635 tokens.

We also include an additional tokenizer (Appendix E.4) for the intervention in Section 6. While a single digit tokenizer would prevent us from evaluating using the same 0-99 periodicity (since only 0-9 would have their own token), we evaluate instead an additional tokenizer that assigns 3 and 4-digit numbers their own token based on a cosine distribution at mode 7. As expected, this change to the tokenizer not only reduces period-10 energy but at the same time increases mode-7 energy of the representations in 1-layer and 8-layer models, while the raw corpus the models were trained on remains the same.

We hope our response has alleviated the reviewer's concerns and would allow them to increase their score. We remain available during the rest of the rebuttal period for any additional questions or concerns.

Thank you for your detailed responses and for addressing my concerns through additional experiments and improvements to the presentation. I appreciate the authors' efforts, and I think the new results strengthen the paper and clarify its contributions. Accordingly, I have increased my score from 2 to 3, with the expectation that all the new experiments, analyses, and revisions described in the rebuttal will be incorporated into the final version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.