acceptodds
Under review as a conference paper at ICLR 2027

Attention Mean Fields Predict Average Representation Dynamics and Illuminate Early Training

Abstract

A language model's representation geometry is not predetermined; it evolves as computation proceeds. Each layer reshapes geometry in part through the effects of attention, which updates each token's representation with information drawn from its context. A faithful account of a model's geometry must capture that dynamic process. To that end, we introduce a mean-field analysis of attention. The average attention a head pays from one token to another defines a kernel that carries representations forward through the model and describes how geometry is transformed. Conditioned on a substantial corpus, the kernel predicts, from the input embeddings and frozen weights alone, the trajectory of each token's mean representation. These predictions are accurate across models spanning GPT-2 to Qwen-3-14B. We show that model-oblivious alternatives, including one built from token co-occurrence, are all poorer predictors. We then develop an attention mean field with more fine-grained conditioning, and demonstrate its value. Specifically, we show that early in training, on real text, replacing every attention head's output with its context-conditional mean field leaves the model's loss unchanged. This effect occurs across three scales of Pythia as well as in OLMo-2-7B. As training progresses through the onset of induction, model behavior and mean-field behavior diverge, with representations becoming contextualized and attention becoming correlated with transported values.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation geometry, language models representations.

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

The authors introduce a mean-field analysis of attention to understand how the geometry of each token's representation evolves. A kernel defined by the average attention a head pays from one token to another describes how token geometry is transformed. The kernel predicts the trajectory of each token's mean representation from input embeddings under fixed weights. The paper shows that early in training, replacing every attention head’s output with its context-conditional mean field leaves the model’s loss unchanged. The observed findings also generalize to large-scale models. As training progresses, the model behavior and mean-field behavior diverge, with representations becoming contextualized.

Critical strengths
  1. The paper is clearly written and easy to follow. The research question of how token representation geometry evolves during initialization and training provides a better understanding of the inner workings of transformer models.

  2. The authors first introduce the contribution $ z =\sum\limits_{j=1}^{q} p_q(x_i) v(x_i) $ (Eq. (1)) that a single head computes for a single context. They then use the mean-field approach to aggregate contributions over a set of contexts that share the same structure.

  3. The generalization of the proposed average representation dynamics across both Pythia and OLMo model families significantly strengthened the paper.

  4. A key strength of the work is its parameter-free mean-field framework, which predicts the evolution of token representations across transformer layers without requiring additional training or fitted parameters.

Critical weaknesses
  1. The mean-field approximation in Section 3 relies on attention–value independence and negligible curvature assumptions. How these assumptions relate to standard transformer models is not clear.

  2. The experiments in Section 5 replace attention only for the most frequent token types, leaving other queries unchanged. Although Appendix L examines larger vocabularies, the results do not establish whether the entire model can be replaced by its mean field.

  3. The authors associate the divergence between model and mean-field behavior with the onset of induction in Section 5. However, the results mainly establish temporal correspondence, without isolating whether induction is responsible for the observed transition.

  4. The paper lacks the anonymous code repo or supplementary zip for reproducing the results provided in the paper.

Questions for authors

Questions:

  1. Section 3.1 defines a token type as an entry in the model's vocabulary. However, the authors could explain what a token type represents from a natural language perspective. In particular, it would be helpful to clarify whether token types correspond to distinct lexical units, how they differ from token positions and grammatical categories (e.g., nouns, pronouns, and adjectives), and how they are identified and grouped across a large corpus. A concrete example illustrating how multiple occurrences of the same token type are handled, despite differences in their positions and linguistic contexts, would help understand the interpretation of token types and the dynamics of their representation across layers.

  2. The authors associate the divergence between model and mean-field behavior with the onset of induction in Section 5. However, the results mainly establish temporal correspondence, without isolating whether induction is responsible for the observed transition. Did the authors observe the induction head behavior?

3 comments
AuthorsOpenReview ↗

Responses to Critical Weaknesses

  1. Relationship of attention-value independence and negligible curvature assumptions to real models. In the paper, we test both of these assumptions directly in standard pretrained models (GPT-2, Gemma-2, Qwen-3-14B, Llama-3). Negligible curvature holds because occurrences of a type cluster tightly around their centroid (Appendix C): the value map at the centroid matches the average value map in direction to within 1 − cos < 3×10⁻⁴ in every layer of both GPT-2 and Gemma-2. A per-layer error decomposition (Appendix G, Fig 15) shows that this approximation costs at most ~0.005 in centered cosine per layer. Attention–value independence holds at the corpus level in trained models: dropping the covariance term likewise costs at most ~0.005 per layer (Fig 15). This is why the mean field predicts average representations well. On the other hand, at the level of individual contexts, independence is exactly what contextual processing violates. A head that attends according to what a particular context’s values carry, rather than according to type, departs from its mean field. The mean field therefore serves as a null model for contextual processing, and §5 uses it in this way.

  2. Replacing the entire model by its mean field. We agree that our experiments do not replace the entire model, and we will say so explicitly. However, the substitution is more complete than "frequent types only" suggests. At every covered position, every head in every layer is replaced, and attention to all key types, including the rarest, is accounted for: untracked keys are pooled into a shared "other" bucket. Only queries whose own type is untracked keep the model's computation (35% of positions at N = 1000). The results also do not depend on where that boundary falls, which we demonstrate in Appendix L. As shown there, raising coverage from 65% to 80% (N = 1000 to 4000, at matched estimation quality) leaves ΔL within 0.001 nats through step 64, and every landmark of the trajectory falls at the same checkpoint. We note that the trends shown in Appendix L across varying N don’t show signs that the early-phase agreement would fail at full coverage. Only the size of ΔL after the departure grows with coverage, as expected if the remaining positions depart too. Full coverage would require kernel rows for rare query types, whose pair statistics cannot be estimated reliably, because support per pair falls quadratically with N (Appendix M).

  3. Is induction responsible for mean-field deviation during training? We do not claim a causal link; suppressing induction-head formation would require retraining the models, which is beyond our scope. We can, however, show that the mean-field departure is specific to induction heads. For each head of Pythia-160m and each frequent query type, we average the head's actual output over all corpus contexts with that query, and likewise its mean-field prediction. The head's mean-field deviation is one minus the cosine between the two averages, averaged over query types. It is near zero when the mean field reproduces the head's type-level behavior. We also score each head two ways. First, we use a standard test of induction: random token sequences, repeated twice. The induction score is the attention a second-copy token pays to the position after its first occurrence. Second, we use a test that detects previous-token behavior: the previous-token score is the attention a token pays to the immediately preceding position.

    We have added this experiment as the new Figure 5 in the updated paper. The figure shows the Spearman correlation across all 144 heads between mean field deviation and each score over training. Through step 256, deviation is unrelated to induction score (all correlations are not significant, ie p ≥ 0.16). The correlation rises to 0.71 at step 1000 and stays between 0.47 and 0.63 thereafter. On the other hand, previous-token heads form in the same window but show the opposite relation (−0.60 at step 1000, −0.39 to −0.55 after). Their deviations rank in the 14th–26th percentile, versus 83rd–94th for induction heads. Deviation thus localizes to induction heads, not to attention reorganization in general. The earliest rise in ΔL (steps 128–256) precedes both head types; we attribute it to the mean field's insensitivity to token order.

  4. Code Repository. Thank you for alerting us to this oversight. An anonymous code repository that reproduces all results of our paper is at https://anonymous.4open.science/r/attention-meanfield-repro-818B/

AuthorsOpenReview ↗

Responses to Questions for Authors

  1. Provide better illustration of how token types relate to lexical units and token positions, and how multiple occurrences of the same token type are handled despite differences in position and context. A token type is an entry in the model’s tokenizer vocabulary. Types therefore do not correspond to lexical units or grammatical categories. A single type can span several senses and parts of speech, and one word can map to several types. For example, in GPT-2, “ May” (with a leading space) is one type, used both for the month (“delayed until May”) and for the surname (“Theresa May”); also “ may” (the modal), and “May” (no leading space, e.g. at the start of a line) are distinct types. An occurrence is a type at a particular position in a particular context. Both occurrences of “ May” above share the type, but each has its own representation at every layer. At the input these representations differ only by position embedding; as attention mixes in context, the month and surname uses diverge. A type’s centroid at layer ℓ is the average of its representations over all its occurrences in the corpus sample (2M tokens), regardless of position or sense, so the centroid of “ May” is a frequency-weighted blend of its uses. The mean field predicts how this blend evolves across layers. Separating the senses within a particular context is contextual processing, which is exactly what the mean field does not capture and what the departure measures of §5 (ΔL, D_MF) quantify. We will add an illustration of this sort to the revision.

  2. Relationship between onset of induction and mean-field deviation. We address this in response 3 to critical weaknesses (above).

Reviewer_vFuPOpenReview ↗

Response to Rebuttal

I thank the authors for their careful response to the review. The authors have responded to all my concerns.

Reviewer_jfKiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a mean-field theory of the propagation of activation centroids---the mean representation in a fixed layer on positions sharing the same token type. By assuming the independence of attention maps and values and the commutativity of the value map with the conditional expectations over tokens with a fixed type, the authors derive a simple equation describing how activation centroids propagate through the layers of the transformer. The propagated centroids stay within mean-centred cosine similarity ≥.75\geq .75 through most layers of 3 out of the 4 tested model classes. Furthermore, the authors develop a more detailed, context-conditional mean-field approximation that retains the token composition of each context. This approximation is shown to capture the early dynamics of Pythia and OLMo models of different sizes.

Critical strengths
  1. The mean-field construction is clearly presented, with its main assumptions stated explicitly. The empirical validation spans four model families and includes comparisons with simpler kernels. These experiments provide evidence that the measured attention kernels capture a significant fraction of the propagation of representation centroids.

  2. The training dynamics experiments provide clear evidence of the validity of the context-conditional approximation at the beginning of training. The consistency across three Pythia sizes and OLMo-2-7B makes it a substantial empirical finding.

Critical weaknesses

In general, the authors did not motivate their study, nor commented on the possible implications. In particular,

  1. The connection between centroid evolution and understanding representation geometry is underdeveloped. Algorithm 1 provides a compact description of average centroid representations, but the paper does not demonstrate which specific geometric properties or transformer behaviours this description helps explain. In other words, the empirical agreement between predicted and measured centroids does not, by itself, make clear what new understanding of representation geometry can be provided by this framework.

  2. The interpretation of the training dynamics needs a clearer connection to existing work on distributional simplicity bias, particularly Rende et al. (2024), which demonstrates the progressive learning of higher-order token interactions. In addition, the paper does not establish what drives the departure from the mean-field regime: the temporal coincidence with induction formation does not identify the mechanism, and stability of the corpus-averaged kernel alone does not establish attention–value independence.

Overall, despite the authors referring to the mean-field analysis as "a useful tool for understanding representation geometry", there is no clear use case emerging from the manuscript.

Questions for authors
  1. Can the authors demonstrate a concrete use of the predicted centroid trajectories to explain a geometric property or transformer behaviour? What insight does the mean-field description (which I understand to require measuring some frozen-weight behaviour anyway) provide beyond directly measuring the centroids at each layer?

  2. How do the early-training loss curves compare with those obtained by replacing attention with uniform attention, or by setting attention outputs to zero, at the same query-token positions? For the uniform-attention control, retaining the same corpus-average value vectors would isolate the contribution of the measured attention kernel. These comparisons would help establish whether the mean-field agreement reflects information captured by that kernel, or whether simpler replacements already preserve early-training performance.

  3. Could the authors clarify the hypothesis that a “low-order corpus-average operator should capture a substantial share of model behavior”? Does “low-order” refer to the kernel being indexed by pairs of token types, or to a restriction on the statistical dependencies captured by the model? What constitutes a “substantial share” of behaviour in this statement? The introduction should distinguish the claims supported by accurate prediction of average centroids from those supported by preservation of early-training next-token loss under the mean-field approximation.

5 comments
AuthorsOpenReview ↗

Response to Critical Weakness 1

We appreciate the referee’s comments on the paper which we have addressed directly and believe will improve the paper’s impact.

Providing explanation of specific properties of representation geometry. MFA provides a new tool for making accurate predictions about how representation geometry evolves. Using MFA, predictions about the evolution of geometry can be made by directly examining mean field attention matrices. This is much more direct and more computationally tractable (see details in our answer 3 to Review #1) than attempting to predict representation geometry from model weights. Given this focus, we have not emphasized results involving specific examples of representation geometry. However, the referee’s suggestion will definitely improve the paper.

As an initial demonstration of what the mean field framework adds to the understanding of representation geometry, we have added a case study on a well-characterized geometry: the circular representation of the months (Engels et al., 2025) which Karkada et al. (2026) explain by translation symmetry (ie, circulant rows) in co-occurrence statistics. That work explains why the circle exists in input embeddings. However, the circle persists across most model layers, and the reasons for its persistence have not been elucidated. We have added our case study of this phenomenon as Section 4.4, Appendix M, and new Figure 19.

Our new results leverage mean field analysis to identify properties of model behavior that are necessary for maintaining the circular structure in GPT-2. The analysis has three results. First, we show that the mean-field rollout (Algorithm 1) of the paper reproduces the month circle layer by layer in GPT-2. That is, the mean-field predicted and true geometries agree at RSA > 0.98 through layer 9, including capturing the sharpening of the circle from the input embeddings that takes place in middle layers. Second, we show that the model’s mean-field attention among months is itself circulant, as Karada et al.’s symmetry suggests: the calendar interval between months explain 71-94% of the variance in month-to-month attention as captured by the mean-field kernel. This is in contrast to under 20% explained variance for random relabelings of the months. Third, because the mean field is an explicit operator on representations (as opposed to the model weights) we can test competing mechanisms for how attention maintains the circle. We can do this by posing counterfactuals to the attention kernel directly. We pose two hypothesized causes a priori: (a) that circulant structure among calendar neighbors, which maps a circle to a circle, is necessary for maintaining circularity; and (b) that self-attention, which adds to each month’s representation a value mapping of itself, maintains circularity.

Our results show that removing circulant structure in attention leaves the predicted geometry unchanged (RSA > 0.999), while removing self-attention degrades the circular geometry progressively across layers of the model (RSA drops from 0.98 to 0.90 by late layers). Thus, although attention displays the symmetry that Karkada et al. predicts, that structure is not what preserves the circle; rather, the circle is preserved by self-attention.

This offers an initial demonstration of the kind of understanding that we believe mean-field attention can bring to the study of geometric properties of representations. Note that direct counterfactual manipulation of attention patterns in the full model is complex to properly pose and implement, and the circulant structure of attention is not necessarily evident directly in model weights. Hence, the use of mean-field analysis is critical to this insight, which could not have been as easily obtained through direct model manipulation.

AuthorsOpenReview ↗

Response to Critical Weakness 2

Connection to distributional simplicity bias. We thank the referee for pointing out the connection to Rende et al. (2024), which also alerted us to Belrose et al. (2024), both of which are relevant and which we will discuss in the revision. Both document a distributional simplicity bias, in which transformers learn the statistics of their data in order of increasing complexity. Belrose et al. show that Pythia first fits unigram and then bigram statistics. Rende et al. build “clones” of a data set that retain token interactions only up to some fixed order. They then show that the loss on low-order “clones” saturates early while higher-order interactions take longer to be learned.

As we will describe in the revision, our results are consistent with this picture and complement it. Rende et al. characterize which statistics of the data a model has learned, by testing it on limited-order interactions. Our work characterizes what the model’s attention computes, by intervening on the model. In the early phase we identify, the model uses its context only through which token types are present, not their order. This is a specific form the simplicity bias takes inside the attention mechanism, during training. Our analysis also helps localize the change; the mean-field kernel is nearly static during the first phase (and only reorganizes at the onset of induction​​). This means that learning in this early stage occurs in the value maps, MLPs, and embeddings rather than in attention routing (§ 5.3). This complements Rende et al.’s observation that layers become active sequentially as interactions of higher order are learned. We will add this discussion to the related work (§ 2) and to § 5.

Why does induction formation drive departure from the mean-field regime? This follows from a property of the mean field that the paper already relies on: it is order-invariant. With each head replaced by its context-conditional mean field, a head's output depends only on the query's type and on which types are present in the context, not on their order. The values it moves are fixed type averages. ΔL is therefore the loss the model avoids through computation that a bag-of-types predictor can’t represent – chiefly the use of token order. The departure marks the onset of that computation. This identifies the class of mechanism, though not a specific circuit. We do not claim that induction alone causes it: ΔL begins to rise around step 256, which is before the presumed induction window (steps 1000–2000), and we will revise §5.2 to present that window as a reference point rather than a cause.

Stability of the corpus-averaged kernel alone does not establish attention–value independence. We agree that kernel stability alone does not establish it, and will revise §5.3. The proper evidence is the substitution itself, which replaces each type's value with its corpus average. If values do not vary with context, Cov_s(A_st, v̄_t) = 0 identically, so the fact that the substitution costs essentially nothing before the departure (ΔL within a few millinats) is direct evidence that Assumption 1 holds in that regime. Kernel stability adds that the attention pattern itself does not change during this period.

AuthorsOpenReview ↗

Responses to Questions to Authors

  1. Demonstrating a concrete use of predicted centroid trajectories. We address this in our response to Critical Weakness 1 above.

  2. Illustrating that mean-field agreement in early training reflects true kernel-specific information. We ran both controls on Pythia-160m at the same query positions and evaluation tokens as Figure 4, with uniform attention over the context using the same corpus-average value vectors. Before the departure (steps ≤ 64), uniform attention matches the model as closely as the mean field does (within 0.001 nats of each other), whereas zeroing attention costs up to 0.17 nats. Early agreement therefore reflects the type-averaged content that attention moves, not structure in the kernel, consistent with §5.3: early in training a head's attention to a type is set by its frequency, which is uniform attention over the tokens present. The kernel becomes informative at the departure: from step 256 through the induction window it recovers 0.07–0.43 nats beyond uniform attention (about 40% of the uniform control's excess loss at steps 256–512).

  3. Clarifying “the hypothesis that a low-order corpus-average operator should capture a substantial share of model behavior.” “Low-order” refers to the operator: the kernel is indexed by pairs of token types (query type, key type), in the same sense as the pairwise statistics (co-occurrence, PMI, bigram) in the work cited. It does not restrict the dependencies the model can capture. “Substantial share” stands for two distinct claims. (i) For average representations in trained models, the corpus kernel predicts each type’s layer-wise centroid with centered cosine above 0.9 through the middle layers and above 0.75 at the last (GPT-2), outperforming co-occurrence and context-free alternatives (§4). (ii) For behavior early in training, replacing attention by its context-conditional mean field at the 65% of positions with tracked query types changes the loss by a few millinats while the loss falls by more than three nats (§5). The two claims are supported by different evidence and are independent. We thank the referee and will sharpen the wording in both the Related Work and the Introduction.

Reviewer_t3NvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper defines a type-to-type attention kernel and uses it to approximate each head's average contribution to token-type centroids. Iterating the resulting operator with frozen model weights predicts layerwise centroid trajectories and outperforms context-free, co-occurrence, and several kernel ablations across models. The authors then construct a context-conditional mean field and replace head outputs during forward passes over checkpoints of three Pythia scales and OLMo-2-7B. Early in training, the replacement has almost no loss cost; around the reported induction-formation window, model behavior, head contributions, and the attention kernel depart from the mean field. The appendices include cross-corpus transfer, top-N sensitivity, error decomposition, and a detailed Llama-3.2-3B rollout failure analysis.

Critical strengths
  1. The paper goes beyond correlational analysis: it replaces attention-head outputs with a context-conditional mean field and measures the causal effect on loss across checkpoints and model scales. This makes the early-training result genuinely informative.

  2. The empirical audit is broad and careful. Multiple architectures, corpora, kernel ablations, top-N sensitivities, and per-layer error decompositions are included. The authors also expose and analyze the Llama-3.2-3B open-loop failure.

Critical weaknesses

See Questions

Questions for authors
  1. If a kernel from an earlier checkpoint, another seed, or another model is frozen and applied to later checkpoints, can it predict the time of departure? This would separate description from genuine dynamics prediction.

  2. Does suppressing or ablating induction-head formation delay the mean-field departure, establishing a causal rather than temporal connection?

  3. What are the compute, memory, and sample complexity of estimating/storing these kernels relative to the behavior they compress?

  4. Is there any clear guidance for improving or designing training procedures?

2 comments
AuthorsOpenReview ↗

Author Response

We thank the reviewer for the careful review and valuable questions.

  1. To clarify, “predicting dynamics” in our title refers to changes in representation geometry across layers of the model, rather than across checkpoints in training. With respect to our training results, we do not claim to forecast when the departure happens. Further, kernels from different seeds or different models are not transferable, because each kernel is specific to a head, and heads are not preserved across models or model seeds. Figure 7 shows how the kernels change across checkpoints.

  2. We do not claim a causal link; suppressing induction-head formation would require retraining the models, which is beyond our scope. We can, however, show that the mean-field departure is specific to induction heads. For each head of Pythia-160m and each frequent query type, we average the head's actual output over all corpus contexts with that query, and likewise its mean-field prediction. The head's mean-field deviation is one minus the cosine between the two averages, averaged over query types. It is near zero when the mean field reproduces the head's type-level behavior. We also score each head on random token sequences, repeated twice. The induction score is the attention a second-copy token pays to the position after its first occurrence. The previous-token score is the attention a token pays to the immediately preceding position.

    Results

    We have added this experiment as the new Figure 5 in the updated paper. The figure shows the Spearman correlation across all 144 heads between mean field deviation and each score over training. Through step 256, deviation is unrelated to induction score (all correlations are not significant, ie p ≥ 0.16). The correlation rises to 0.71 at step 1000 and stays between 0.47 and 0.63 thereafter. On the other hand, previous-token heads form in the same window but show the opposite relation (−0.60 at step 1000, −0.39 to −0.55 after). Their deviations rank in the 14th–26th percentile, versus 83rd–94th for induction heads. Deviation thus localizes to induction heads, not to attention reorganization in general.

  3. Thank you for asking us to clarify this important point. Memory complexity: the kernel stores one N×N table per head (N = 1000 types), about 10⁶ entries per head: 577 MB for GPT-2 small, 834 MB for Gemma-2-2B, and roughly 10–13% of parameter count for the 8–14B models. The raw attention it summarizes on our 2M-token sample would occupy about 300 GB, in other words, roughly 500× more space. Note as well that the kernel transfers to unseen text and other corpora (Appendix F), so does not need to be recomputed in those settings. Sample complexity depends on vocabulary size rather than model size. At N = 1000, 2M tokens suffice: as mentioned, estimates transfer across corpora and are unchanged at 40M tokens (Appendix M). Note that 2M tokens is about 0.001% of Pythia's training data. The required sample grows roughly as N², as expected for Zipfian type frequencies, which Appendix M measures directly (4× at N = 2000, 15–20× at N = 4000). Compute complexity: Estimation of all kernels (simultaneously) requires one gradient-free forward pass over the sample, plus accumulating each attention row into its query type's row. This costs about as much as inference. On 2M tokens that comes to about 1–2 hours on a laptop for the small models.

  4. This is a fair question, but basically lies outside the scope of the current work, which does not give direct guidance on training procedures. Rather, mean field analysis offers a tool for understanding model behavior (by separating the part of a model’s computation that a corpus-average, type-level kernel can express from the part that depends on the specific context). That said, some of our results do bear on how training proceeds. First: early in training, loss falls by more than three nats while the attention kernel is essentially unchanged. This shows that learning in this phase occurs in the values, MLPs, and embeddings rather than in attention routing. Further, the departure from the mean field occurs at roughly the same number of training tokens across a wide range (75x) of model sizes. Finally, the mean-field loss levels off at a similar value across models, so the improvement from model scale (across the 75x size range) is seen to arise from contextual computation. We believe that these observations could inform the design of training interventions, although testing any such intervention is beyond the scope of the paper. In the revision we will state this scope explicitly in the conclusions.

Dear Reviewer t3Nv,

Could you please review the authors’ rebuttals and follow up with any further comments or questions?

Many thanks for your time and efforts.

Best regards, AC

Comments on the paper

1 comment

Dear Reviewers,

As the AC–reviewer discussion phase is about to begin, please highlight any disagreements or inconsistencies across the reviews, as well as any concerns that remain unresolved after the author rebuttal.

Please share your perspectives on these points and engage with one another to help us reach a well-supported assessment of the paper.

Thank you for your time and contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.