acceptodds
Under review as a conference paper at ICLR 2027

Opinion Leader Dynamics: How Sparse Attention Shapes Token Clustering

11 views

Abstract

Sparse attention reduces the quadratic cost of global self-attention while retaining strong empirical performance, but how its restricted interactions shape the evolution of token representations remains theoretically underexplored. Modeling tokens as particles on the unit sphere, we introduce **opinion leader dynamics**, a framework that identifies two mechanisms through which token groups converge internally while maintaining distinct limiting directions. In the explicit model, fixed representatives induce a potential that attracts tokens toward distinct local maxima. In the implicit model, disconnected interaction groups evolve toward separate consensus directions. We formulate both models as *reverse Wasserstein gradient flows* and establish exponential convergence under suitable conditions. We further connect these theoretical predictions to token evolution in frontier sparse-attention LLMs that motivate our framework. Across four benchmarks, \\textcolor{#5178a1}{\\textbf{Kimi-K3}}, \\textcolor{#b11f23}{\\textbf{MiniMax-M3}}, and \\textcolor{#5178a1}{\\textbf{DeepSeek-}}\\textcolor{#b11f23}{\\textbf{V4-Flash}} consistently exhibit clearer cluster separation and higher clustering scores than the dense-attention model \\textcolor{#148e6f}{\\textbf{GLM-4.7-Flash}} in projected token representations. These observations support the relevance of the predicted multiple-group structure to trained frontier LLMs, while finite-particle simulations illustrate the theoretical convergence behavior. Together, our results connect restricted token interactions to distinct group-level attractors, providing a dynamical account of how sparse attention can support alignment within groups while preserving separation between them.

Primary area:
learning theory
Keywords:
Token Clustering, Attention Dynamics, Interacting Particle Systems, Sparse Attention

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hZkmOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper considers a dynamic model for sparse attention, analyzing the dynamics of token interactions throughout depth. The main phenomenon that the authors try to capture is the clustering of tokens through the dynamics. In the analysis, they consider two models: an explicit model in which the tokens interact with fixed representatives, and an implicit model in which tokens interact with one another under sparsity constraints. The authors prove that, under specific assumptions and constraints, the tokens cluster onto clusters over limiting directions. They show that these results also hold in practice by using particle dynamics simulations in settings that fit their model exactly, and also by analyzing the hidden representation of LLMs at the final layer.

Critical strengths
  • The theoretical analysis is comprehensive and well written, with a high level of mathematical rigor. The authors carefully develop the convergence arguments and state the assumptions under which their guarantees hold.

  • The paper complements its theoretical results with useful numerical simulations and empirical comparisons with actual pretrained models and additional ones that fit their analyzed setting exactly; In both they show the clustering happens in practice. This effort to connect the theory with practical observations is valuable and goes beyond what is commonly included in theoretical work.

Critical weaknesses
  • The paper aims to explain how sparse interactions support separated token groups. However, in the implicit model (which is the practical one in my view), separation relies on the disconnected interactions assumption and the disjoint caps assumption. Accordingly, the theorem establishes the limiting dynamics within already separated groups while simply preserving their separation. This makes the result not so surprising.

  • My main concern is the significance of the contribution beyond existing work. References [1] and [2] study closely related mechanisms and dynamics and attraction to frozen centers. Although these works are cited, the paper should state precisely which assumptions, conclusions, or quantitative bounds are new and explain the importance of these advances. Beyond establishing clustering under the stated assumptions, stronger insights into the meaning or practical consequences of the resulting clusters would strengthen the contribution.

[1] Jingkun Liu, Yisong Yue, Max Welling, and Yue Song. Krause synchronization transformers. ICML, 2026. URL https://arxiv.org/abs/2602.11534.

[2] Nikita Karagodin, Yury Polyanskiy, and Philippe Rigollet. Clustering in causal attention masking. NeurIPS, 2024. URL https://arxiv.org/abs/2411.04990.

Questions for authors
  • Could you provide additional intuition about the explicit model? What do the fixed representatives correspond to in practical architectures, and why is it reasonable to keep them fixed throughout the dynamics?

  • What is the practical interpretation of the resulting clusters? Does this clustering have implications for the expressivity of attention or the information retained in token representations? What practical insights or predictions follow from the analysis?

  • The evaluated LLMs include feed-forward layers and other components absent from the theoretical model. What evidence indicates that the observed clustering is driven by the attention dynamics analyzed here?

  • Could you provide intuition on how Softmax normalization and multiple attention heads might affect the clustering behavior? Which conclusions are expected to remain valid, and which parts of the analysis would require modification?

Clarifications

  • Corollary 2.1 is presented without a connection to a specific theorem. Please identify in the main text the theorem that Corollary 2.1 is built on.

  • Assumption 2.2 is difficult to follow. Please distinguish its concentration, separation, and representative-dominance conditions explicitly and explain the roles of δ\delta, Δ\Delta, η\eta, and α\alpha. Including the definitions of Ia(α)I_a(\alpha) and Ha(α)H_a(\alpha) alongside the assumption would also improve readability.

  • Please state in the caption of Figure 4 which architectural interpretation is associated with each model: explicit, implicit, hybrid, or the dense-attention baseline.

6 comments
AuthorsOpenReview ↗

Response to Reviewer hZkm (Part 1)

We thank Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}} for positive assessment of our theoretical analysis and empirical investigation. We believe it would be always helpful to connect theory with pratical experiments nowadays. Below, we clarify the contributions of both theoretical models, their relationship to prior work and practical architectures, and the interpretation of our experiments. We also discuss Softmax normalization and multiple attention heads, and address the requested presentation improvements.


1. Contribution of the implicit\textcolor{#b11f23}{\textbf{implicit}} model (W1)

We agree that Assumption 2.4 excludes cross-group interactions and that Assumption 2.5 places the groups in disjoint initial caps. Theorem 2.6 establishes cap invariance, ensuring that this separation persists. Its additional contributions concern the dynamics within each sparsely connected component:

  • Exponential consensus. All tokens converge exponentially to a common direction despite interacting through sparse, overlapping neighborhoods. This consensus direction depends on their initial states and interactions, but is not prescribed in advance.
  • Quantitative dependence on connectivity. The convergence-rate bound explicitly depends on the component’s normalized Laplacian spectral gap. For fixed cap radius and inverse temperature, the guaranteed rate scales as nk−2n_k^{-2} on a self-looped path of nkn_k tokens and remains bounded away from zero for graph families with a uniformly positive spectral gap.

We note that group separation alone does not establish either within-group consensus or these quantitative convergence rates. Our result characterizes how sparse connectivity governs alignment within groups while their separation is preserved.


2. Contribution beyond existing work (W2)

Comparison with Krause Attention [1]. Krause Attention is closely related to our implicit\textcolor{#b11f23}{\textbf{implicit}} model. Its clustering analysis establishes exponential convergence under a small-cap condition. Our analysis allows a wider initial cap and proves cap invariance and exponential consensus on fixed sparse interaction graphs. Our explicit convergence-rate bound quantifies the effect of within-group connectivity through the normalized Laplacian spectral gap.

Comparison with Karagodin et al. [2]. In their frozen-token analysis on the circle, Karagodin et al. [2] study causal self-attention with additional attraction to fixed tokens, which idealize quasi-stationary cluster centers. Our explicit\textcolor{#5178a1}{\textbf{explicit}} model is motivated by representative-based information compression in practical sparse attention architectures: each initial token group is summarized by a fixed representative, and tokens on Sd−1\mathbb S^{d-1} interact exclusively with these representatives. These representatives summarize token groups, reflecting compression mechanisms in practical sparse attention. We further go beyond convergence guarantees to derive explicit convergence-rate bounds and quantitative lower bounds on the pairwise separation of the limiting clusters. We have added this discussion in Appendix B (Comparison with Prior Clustering Analyses) of the revised manuscript.


[1] Liu, J., Yue, Y., Welling, M., and Song, Y. Krause synchronization transformers. ICML, 2026.

[2] Karagodin, N., Polyanskiy, Y., and Rigollet, P. Clustering in Causal Attention Masking. NeurIPS, 2024.

AuthorsOpenReview ↗

Response to Reviewer hZkm (Part 2)


3. Motivation and scope of the explicit\textcolor{#5178a1}{\textbf{explicit}} model (Q1)

Architectural motivation and intuition. Practical sparse attention architectures based on landmark representations [3, 4] or MLP-compressed representatives [5] summarize token groups and use these summaries to aggregate information or guide retrieval. Such representatives can provide common reference directions towards which tokens evolve. Our explicit\textcolor{#5178a1}{\textbf{explicit}} model abstracts this operation by constructing representatives from initial token groups and studying how they guide subsequent token evolution.

The representatives jointly induce an attraction potential. Its local maxima determine the attracting directions, which need not coincide exactly with individual representative directions. This provides a mathematical description of how compressed summaries can guide groups of tokens toward distinct representations.

Fixed representatives. Fixed representatives are the central assumption of our model and is essential to our current analysis. It produces a time-independent attraction potential, allowing us to relate its geometry to clustering structure and convergence rates. This is an idealization of the practical architectures, where representatives evolve across layers. Studying attention dynamics with evolving representatives is therefore an important direction for our future work. We have clarified this assumption and the corresponding direction for future work in Appendix A of the revised manuscript.


[3] Mohtashami, A., and Jaggi, M. Landmark attention: Random-access infinite context length for transformers. NeurIPS, 2023.

[4] Leng, J., Hu, X., Wang, J., Li, J., Wu, W., and Lu, Y. Understanding and improving length generalization in hierarchical sparse attention models. ICLR, 2026.

[5] Yuan, J., Gao, H., Dai, D., Luo, J., Zhao, L., Zhang, Z., ... and Zeng, W. Native sparse attention: Hardware-aligned and natively trainable sparse attention. ACL, 2025.


4. Semantic and functional structure of token clusters (Q2)

We thank the reviewer for this thoughtful question, which prompted us to examine what the observed final-layer clusters represent. We analyzed the tokens assigned to each cluster in DeepSeek-V4-Flash across 100 samples each from ARC-Easy and MATH.

The clusters exhibit clear semantic and functional coherence, grouping tokens associated with answer choices, mathematical notation, and prompt templates. These findings give the geometric clusters a practical interpretation: they correspond to recognizable aspects of token content and function across tasks. We report this new analysis in Section 3.4 (Cluster Semantics) of the revised manuscript.

Our theoretical results show how within-group alignment can coexist with distinct group-level representations. The semantic analysis further provides empirical evidence that such groups can have interpretable content in trained LLMs! Quantifying the implications for expressivity or information retention remains an important direction for further study.


5. Token clustering before and after the final FFN (Q3)

To examine the contribution of the final FFN, we compute the cosine silhouette scores across 100 random ARC-E samples before and after the final layer FFN in DeepSeek-V4-Flash. The scores are similar: 0.82±0.080.82 \pm 0.08 before and 0.85±0.120.85 \pm 0.12 after the FFN. These results show that substantial cluster separation is already present before the final FFN and that the scores remain similar in magnitude afterward. We have added the analysis in Appendix G.4.5 and regard a fuller investigation of how individual components shape token dynamics as future work.

AuthorsOpenReview ↗

Response to Reviewer hZkm (Part 3)


6. Clustering under softmax normalization and multi-head attention (Q4)

Softmax normalization. Introducing softmax normalization affects our two models differently:

  • Explicit model. Let F(x)=∑aρaeβ⟨x,za⟩Px⊥(za), Z(x)=∑aρaeβ⟨x,za⟩F(x)=\sum_a\rho_a e^{\beta\langle x,z_a\rangle}P_x^\perp(z_a),\ Z(x)=\sum_a\rho_a e^{\beta\langle x,z_a\rangle}. The normalized dynamics are x˙=F(x)/Z(x)\dot{x}=F(x)/Z(x). For fixed representatives, Z(x)Z(x) is bounded above and away from zero on the sphere. Dividing the vector field by Z(x)Z(x) changes only the speed at which each token moves along its trajectory. The equilibria, attraction basins, and limiting clusters are therefore preserved. Exponential convergence persists, with adjusted rate bounds.
  • Implicit model. Under the same fixed-graph and cap assumptions, normalization preserves the interaction support and uniformly positive weights on existing edges. Cap invariance, componentwise consensus, and separation between components therefore persist, preserving the qualitative multiple-cluster conclusion. However, the convergence-rate bound requires further analysis, as the token-dependent normalization breaks symmetry under fixed degree weights.

Multiple attention heads. Considering multiple attention heads also has distinct effects on clustering in our two models:

  • Explicit model. Combining the head vector fields with positive weights yields the spherical gradient of the potential Ψ(x)=∑h=1Hchβh∑a=1mhρh,aeβh⟨x,zh,a⟩,\Psi(x)=\sum_{h=1}^{H}\frac{c_h}{\beta_h}\sum_{a=1}^{m_h}\rho_{h,a}e^{\beta_h\langle x,z_{h,a}\rangle}, where HH is the number of heads, ch>0c_h>0 is the weight of head hh, and βh>0\beta_h>0 is its inverse temperature. Head hh has mhm_h fixed representatives zh,az_{h,a} with positive weights satisfying ∑aρh,a=1\sum_a\rho_{h,a}=1. Combining heads can shift the attractors, but the gradient structure is preserved. If the combined field preserves disjoint caps and Ψ\Psi is uniformly strictly geodesically concave within each cap, the same proof strategy yields exponential convergence to distinct attracting directions. The dominance conditions and rate bounds must therefore be established for the combined potential in the multi-head settings.
  • Implicit model. With positive head weights, the combined interaction graph is the union of the individual head graphs. If all heads respect a common group partition and the union graph is connected within each group, interactions remain confined to groups while allowing information to propagate throughout each one. Under the cap assumptions, componentwise consensus and distinct group limits are preserved. Quantitative rates require analysis of the combined interaction operator.

7. Clarification

Corollary. Thank you for identifying the missing connection. Corollary 2.1 follows the classical Łojasiewicz convergence theorem for analytic gradient flows [6, Section 2]. The potential is real analytic and the sphere is compact, ensuring that each trajectory converges to a critical point. The underlying argument is provided in Appendix F.1.2. We have added an explicit statement before Corollary 2.1 to discuss this connection.

Assumption 2.2. In the revised manuscript, we have reorganized the assumption into three conditions and clarified their roles:

  • Concentration: δ\delta bounds the initial spread of each group, while η\eta bounds the angle between its center and representative direction. The condition δ+η<α\delta+\eta<\alpha places each group inside a cap of radius α\alpha around its representative.
  • Separation: Δ\Delta is a lower bound on the distance between group centers. The condition 2(α+η)<Δ2(\alpha+\eta)<\Delta ensures that the representative-centered caps are disjoint.
  • Representative dominance: Ia(α)>0I_a(\alpha)>0 ensures that the vector field points into the cap at its boundary, making the cap forward invariant. Ha(α)>0H_a(\alpha)>0 ensures strict concavity of the potential within the cap, yielding a unique attractor and an explicit exponential convergence rate.

We have placed the definitions of Ia(α)I_a(\alpha) and Ha(α)H_a(\alpha) alongside the assumption to improve readability. Please refer to Section 2.1 of the revised manuscript.

Figure.4. The original manuscript used different colors to indicate the architectural categories. We have now added explicit descriptions in the caption identifying the attention mechanism associated with each model as explicit\textcolor{#5178a1}{\textbf{explicit}}, implicit\textcolor{#b11f23}{\textbf{implicit}}, $\textcolor{#5178a1}{\textbf{hyb}}$$\textcolor{#b11f23}{\textbf{rid}}$, or standard\textcolor{#148e6f}{\textbf{standard}} architectures.


[6] Absil, P. A., Mahony, R., and Andrews, B. Convergence of the iterates of descent methods for analytic cost functions. SIAM Journal on Optimization, 16(2), 531-547.


We thank the reviewer again for the constructive feedback. We hope these clarifications and additional analyses address the concerns, and we welcome further questions and discussion.

Reviewer_bee6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper theoretically and experimentally studies how tokens cluster in standard and sparse attention. For this, tokens are modeled as particles on the unit sphere. The opinion leader dynamics framework (where representations of tokens are probed whose interactions are restricted) is introduced to formalize how clusters are formed. Following the theoretical grounding, experimental evaluation shows that a dense-attention model has a lower cluster separation than the three evaluated sparse attention models.

Critical strengths
  1. The paper presents an interesting theoretical perspective that helps better understanding the effects on sparse attention mechanisms. For me, this is an interesting perspective that I was not aware of before, making it an interesting read.
  2. The experimental evaluation provides evidence for analogous behavior in modern LLMs. The theoretical framework allows to ground the observation in a principled formulation.
Critical weaknesses
  1. While the method is presented well and the relation to modern LLMs is provided (with experimental evidence), the underlying motivation for having such convergence guarantees is not sufficiently explained. This could be attributed to the fact that I am not very familiar with the works in this field. Therefore, I would appreciate a concise summary in what way the paper’s findings can be interesting and relevant to the community. If the authors find a convincing argument for this, it would substantially improve my current view on the paper.
  2. The framework simplifies real transformers but the effects of this are only discussed in a limited way. It would be important to explain in what way conclusions might be potentially limited due to this simplification step.
  3. Further, the discussion of the relations between sparse and dense LLMs remains limited. While the appendix assigns different attention types in modern LLMs to the two types, a discussion of it might be actually main paper material, to improve the depth of the discussion and relation to current architectures.
Questions for authors
  1. What would happen if the scores were not computed in the reduced representation space? Does it have an impact to compute it in the UMAP space? How sensitive are the results do the dimensionality?
  2. How do evaluations change if not the last layer but a different layer is selected?
5 comments
AuthorsOpenReview ↗

Response to Reviewer bee6 (Part 1)

We thank Reviewer bee6\textcolor{blue}{\textbf{Reviewer bee6}} for the thoughtful review and for finding our theoretical perspective interesting. Below, we clarify the motivation and practical significance of the framework, the scope of its guarantees, its relevance to modern LLMs, and the clustering evaluation.


1. Why our framework matters (W1)

Motivation and significance of the convergence guarantees. Sparse attention reduces the quadratic cost of global self-attention while retaining strong empirical performance, but how its restricted interactions shape the evolution of token representations remains theoretically underexplored. Our framework studies when these interactions allow tokens to align within groups while maintaining distinct representations across groups.

The convergence guarantees make this description quantitative: they identify sufficient conditions for persistent group structure and characterize how representative geometry and within-group connectivity govern the speed of alignment. This provides a basis for understanding how different sparse attention mechanisms organize information as representations evolve.

Architectural categorization and design implications. Prior analyses have characterized clustering in full and causal standard attention under suitable assumptions [1–7]. Our opinion leader dynamics framework studies sparse attention through two recurring mechanisms: explicit\textcolor{#5178a1}{\textbf{explicit}} mechanisms aggregate information into representatives, while implicit\textcolor{#b11f23}{\textbf{implicit}} mechanisms restrict interactions among tokens. $\textcolor{#5178a1}{\textbf{Hyb}}$$\textcolor{#b11f23}{\textbf{rid}}$ designs combine both mechanisms. To the best of our knowledge, this is the first systematic categorization of attention mechanisms from the perspective of token dynamics.

Applying this categorization to practical architectures highlights the growing role of information compression alongside sparse token interactions. It makes two complementary design choices explicit: which token interactions to retain and how to summarize token information into representatives. For example, hybrid designs can combine compact representations of broader context with selective access to fine-grained token information. These distinctions provide structured bases for comparing existing mechanisms and identifying ways to combine or refine them. We have added this discussion to Appendix A (Trends and Design Implications) of the revised manuscript.

Practical interpretation of the clusters. The original manuscript compared three sparse-attention LLMs with a dense-attention baseline and observed differences in token clustering consistent with the mechanisms studied in our framework. During the rebuttal, we extended this analysis to examine what the final-layer clusters represent in LLMs. Specifically, we examine clusters in DeepSeek-V4-Flash across 100 samples each from ARC-Easy and MATH. The clusters exhibit clear semantic and functional coherence, grouping tokens associated with answer choices, mathematical notation, and prompt templates. These findings give the observed geometric clusters an interpretable role in organizing token content across tasks. The new analysis appears in Section 3.4 (Cluster Semantics) of the revised manuscript.

Finally, as Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}} noted among the paper’s strengths, our effort to connect theory with practical observations “goes beyond what is commonly included in theoretical work.” This connection is central to the framework’s value: it combines a quantitative analysis of sparse attention dynamics with empirical evidence of meaningful organization in the token representations of trained LLMs.


[1] Chen, S., Lin, Z., Polyanskiy, Y., and Rigollet, P. Quantitative Clustering in Mean-Field Transformer Models. arXiv:2504.14697, 2025.

[2] Geshkovski, B., Letrouit, C., Polyanskiy, Y., and Rigollet, P. A Mathematical Perspective on Transformers. Bulletin (New Series) of the American Mathematical Society, 2025.

[3] Karagodin, N., Polyanskiy, Y., and Rigollet, P. Clustering in Causal Attention Masking. NeurIPS, 2024.

[4] Geshkovski, B., Letrouit, C., Polyanskiy, Y., and Rigollet, P. The emergence of clusters in self-attention dynamics. NeurIPS, 2023.

[5] Abella, Á. R., Silvestre, J. P., and Tabuada, P. Consensus is all you get: The role of attention in transformers. ICML, 2025.

[6] Bruno, G., Pasqualotto, F., and Agazzi, A. Emergence of Meta-Stable Clustering in Mean-Field Transformer Models. ICLR, 2025.

[7] Rigollet, P. The mean-field dynamics of transformers. ICM, 2026.

AuthorsOpenReview ↗

Response to Reviewer bee6 (Part 2)


2. Theoretical scope and relevance to modern LLMs (W2 and W3)

Theoretical scope. We agree that the consequences of our simplifications should be stated clearly. Our guarantees concern idealized attention dynamics under specified conditions. Their correspondence with trained LLMs is therefore at the level of mechanisms and observed qualitative behavior.

Roles and limitations of the assumptions. Our assumptions specify sufficient conditions for the guarantees:

  • Assumption 2.2: representative separation and dominance. These conditions ensure a unique attracting equilibrium in each prescribed cap and exponential convergence for tokens initialized there. If representatives are closer or dominance is weaker, the number and locations of attractors may change, so the guarantee of a distinct attracting cluster associated with each representative no longer applies.
  • Assumption 2.4: disconnected interaction groups. Disconnection allows each component to evolve independently toward its own consensus. Cross-group edges couple the dynamics and can alter long-time separation, requiring additional analysis on the convergence.
  • Assumption 2.5: concentrated and separated initialization. Concentration within each cap enables our contraction estimate, while separation between caps ensures distinct limiting clusters. Relaxing cap separation alone preserves the within-component consensus guarantee, but the limiting directions of different components may coincide.

We have expanded the discussion of these assumptions in Appendix A (Justification of Assumptions).

Categorization of modern LLMs. Following the reviewer’s suggestion, we have added a discussion in Section 3.1 of the main paper, categorizing the attention mechanisms used in modern LLMs as explicit\textcolor{#5178a1}{\textbf{explicit}}, implicit\textcolor{#b11f23}{\textbf{implicit}}, or $\textcolor{#5178a1}{\textbf{hyb}}$$\textcolor{#b11f23}{\textbf{rid}}$, with standard\textcolor{#148e6f}{\textbf{standard}} dense attention included for reference. This makes the architectural relevance of the framework directly accessible in the main text. The more detailed categorization of individual attention mechanisms remains in Appendix D.


3. Clustering across embedding spaces, dimensions, and layers (Q1 and Q2)

Clustering in the original space. A major practical challenge is the 16,384-dimensional representation space used in our DeepSeek-V4-Flash analysis. In very high-dimensional spaces, distance concentration can reduce the contrast between nearby and distant points, making local density estimation and density-based cluster identification more difficult. In our experiments, applying HDBSCAN directly to the original representations frequently produced fewer than two non-noise clusters, preventing meaningful computation of the corresponding silhouette scores. This outcome makes the direct HDBSCAN evaluation inconclusive under our current setup.

Choice of UMAP. We use UMAP before HDBSCAN to facilitate clustering in a reduced representation space. This pipeline has precedent in work such as BERTopic, which clusters Transformer-based document embeddings [8].

Sensitivity to the reduced dimensions. The original submission includes an analysis of the sensitivity of clustering scores to the reduced embedding dimension in Appendix G.4.1. For DeepSeek-V4-Flash on ARC-E, scores remain stable in 0.85–0.86 across dimensions of 16, 32, 48, and 64. This stability suggests that clustering scores are stable and robust across the tested dimensions.

Clustering across layers. During the rebuttal, we added a layerwise analysis of clustering scores in 4 frontier LLMs using ARC-Easy samples. The scores generally increase with depth in the evaluated sparse-attention models, while the dense-attention baseline remains close to the Gaussian reference. These results complement the final-layer comparisons by showing how the measured cluster separation evolves across depth. They are qualitatively consistent with stronger group separation in the evaluated sparse-attention models. The analysis appears in Appendix G.4.2 of the revised manuscript.


[8] Grootendorst, Maarten. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv preprint arXiv:2203.05794, 2022.


We thank the reviewer again for the constructive feedback, which has helped us communicate the practical significance and limitations of our framework more clearly. We welcome further questions and discussion.

Reviewer_bee6OpenReview ↗

Follow-up

Thank you for your detailed rebuttal and sorry for the late reply. I am still concerned about whether clustering in reduced space actually captures the structure. Could you please concisely summarize here in what way this reduction does not shape the space in a way that groups just emerge, specifically in the discussed context? Is there maybe a different way to show these structural properties? Could you extend the analysis to higher dimensions (if still feasible)? I am not requiring this but I would appreciate a clearer discussion here (and the paper).

Further, after reading the other reviews, I would encourage clarifying the first concern of reviewer wVGc.

Thanks a lot.

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

The paper extends previous works on token dynamics to another class of Transformer models, focusing on the class that restricts token interaction via sparse attention matrices. The authors propose a framework called opinion leader dynamics, separating two cases, whether the attention explicitly chooses a landmark or just restricts tokens into groups, contrasting them with standard attention. The framework formulates attentions as reverse Wasserstein gradient flows and establishes exponential convergence rates. Simulations and experiments on largescale LLMs in these two cases show stronger multiple-cluster convergence tendency than standard attention.

Critical strengths
  • The paper extends the study of token dynamics to an underexplored family of attentions that restrict interaction between tokens. The authors established a convergence result with a converging rate.
Critical weaknesses
  • W1. It seems the real focus of this paper is to study the case where tokens are disconnected into groups rather than sparsity in general as suggested from the title, abstract, and introduction. In particular, the implicit leader case doesn't support sliding window attention (which is sparse but connected, excluded by Assumption 2.4). Moreover, the assumptions (2.2, 2.4, and 2.5) are not justified to see if they are applicable to the practical LLMs.
  • W2. The formulation of attention is simplified and not practical:
    • The query, key, and value projection matrices are set to identity matrices or a multiple of identity.
    • The attention scores (in Eq (1) and (2)) are not ordinary softmax-normalized attention, which is the standard for practical transformers, and are applied in previous works, eg, [1,2]

Reference:

[1] Abella et al. Consensus is all you get: the role of attention in transformers. ICML 2025.

[2] Karagodin et al. Clustering in Causal Attention Masking. NeurIPS 2024.

Questions for authors
  • Q1. How would the result be extended if the simplifications in W2 are relaxed?
  • Q2. The clustering scores in Table 1 are computed after using UMAP to reduce dimensionality. This could lose information about the original structures or create artifacts that wrongly affect the score. Could the authors provide clustering scores of the original embeddings without using UMAP?
5 comments
AuthorsOpenReview ↗

Response to Reviewer wVGc (Part 1)

We thank Reviewer wVGc\textcolor{red}{\textbf{Reviewer wVGc}} for the thoughtful feedback. We address the scope of the theory, the attention formulation, and the clustering evaluation below.


1. Scope of the theory and justification of assumptions (W1)

Implicit\textcolor{#b11f23}{\textbf{Implicit}} Model & Sliding-Window Attention. Our implicit model captures sparse, connected interactions within each group, including overlapping radius-based neighborhoods that already follow the same locality principle of sliding-window attention. As the reviewer notes, however, our multiple-cluster theorem does not cover a sliding-window graph that connects the entire sequence, because Assumption 2.4 requires disconnected components.

The theorem characterizes consensus within each sparse component and the preservation of separation between components under the stated initialization conditions. Its quantitative contribution is an explicit convergence-rate bound linking within-group connectivity to alignment speed. Allowing cross-group communication would couple the component dynamics and could alter their long-time separation, requiring additional analysis on the convergence. The disconnected setting simplifies the analysis and enables quantitative bounds linking local connectivity to convergence rates.

Motivation and roles of the assumptions. The assumptions used in our framework combine architectural motivations with geometric conditions of prior theory studies that make the analysis tractable:

  • Assumption 2.2 imposes separation and representative-dominance conditions that allow us to characterize distinct attracting regions. The representative-based aggregation and selective retrieval in practical sparse attention [1, 2] motivate this setting.
  • Assumption 2.4 retains sparse, connected interactions within each component while excluding interactions between components. The within-component neighborhoods reflect the locality principle of sliding-window attention [3, 4], while disconnection is an analytical idealization that permits independent groupwise analysis.
  • Assumption 2.5 places each component’s initial states in a concentrated spherical cap and separates the caps of different components, following related geometric conditions in prior attention analyses [5, 6]. This assumption enables us to prove cap invariance and quantify within-group contraction while preserving between-group separation.

These conditions define sufficient assumptions for our theoretical results. We have clarified their motivations and roles in Appendix A (Justification of Assumptions) of the revised manuscript.


[1] Mohtashami, A., and Jaggi, M. Landmark attention: Random-access infinite context length for transformers. NeurIPS, 2023.

[2] Yuan, J., Gao, H., Dai, D., Luo, J., Zhao, L., Zhang, Z., ... and Zeng, W. Native sparse attention: Hardware-aligned and natively trainable sparse attention. ACL, 2025.

[3] Beltagy, I., Peters, M. E., and Cohan, A. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020.

[4] Zaheer, M., Guruganesh, G., Dubey, K. A., Ainslie, J., Alberti, C., Ontanon, S., ... and Ahmed, A. Big bird: Transformers for longer sequences. NeurIPS, 2020.

[5] Geshkovski, B., Koubbi, H., Polyanskiy, Y., and Rigollet, P. Dynamic metastability in the self-attention model. arXiv preprint arXiv:2410.06833, 2024.

[6] Chen, S., Lin, Z., Polyanskiy, Y., and Rigollet, P. Quantitative Clustering in Mean-Field Transformer Models. arXiv:2504.14697, 2025.

AuthorsOpenReview ↗

Response to Reviewer wVGc (Part 2)


2. Attention formulation and possible extensions (W2 and Q1)

Q/K/V Projections. Identity or isotropic projections are established idealizations in analyses of attention dynamics [1, 2, 5]. Here, they allow us to isolate how representative geometry and sparse interaction structure affect clustering.

Extending our convergence-rate and cluster-separation guarantees to general projections requires further analysis. General query/key projections can introduce asymmetric interactions, while value projections can change the directions of token updates. Consequently, the potential structure and contraction arguments used in our proofs need not carry over directly. Related challenges are discussed in [2, Section 9]. Identifying appropriate structural conditions under which our guarantees extend is an important direction for future work.

Softmax normalization. The reviewer correctly and sharply observes that Eqs. (1) and (2) use unnormalized self-attention (USA) instead of softmax-normalized self-attention (SA). We should have stated this distinction more explicitly. For example, the standard attention dynamics defined in [1] are x˙i=Pxi⊥(1n∑jeβ⟨xi,xj⟩xj)\dot{x}_{i}=P_{x_{i}}^{\perp}\left(\frac{1}{n}\sum_{j}e^{\beta\langle x_{i},x_{j}\rangle}x_{j}\right). Our explicit\textcolor{#5178a1}{\textbf{explicit}} model replaces the uniform empirical measure 1n∑j=1nδxj(t)\frac{1}{n}\sum_{j=1}^{n}\delta_{x_j(t)} with the representative measure ∑aρaδza\sum_a\rho_a\delta_{z_a} (∑aρa=1\sum_a\rho_a=1), while our implicit\textcolor{#b11f23}{\textbf{implicit}} model replaces it with the neighborhood empirical measure ∑j∈Ni(R)1di(R)δxj(t)\sum_{j\in\mathcal N_i^{(R)}}\frac{1}{d_i^{(R)}}\delta_{x_j(t)} (∑j∈Ni(R)1di(R)=1\sum_{j\in\mathcal N_i^{(R)}}\frac{1}{d_i^{(R)}}=1). These weights normalize the underlying interaction measures instead of the exponential attention scores.

Actually, both SA and USA are commonly used established frameworks in the literature [1-7]. For example, [1, 2] analyze USA using its reverse Wasserstein gradient-flow structure, while [3-5] study softmax-normalized attention dynamics. Reference [6] formulates mean-field equations for both models, and [7] provides an overview of both formulations. We adopt USA because its reverse Wasserstein gradient-flow structure facilitates the quantitative analysis. The effect of introducing Softmax normalization differs between our two models:

  • Explicit model. Let F(x)=∑aρaeβ⟨x,za⟩Px⊥(za), Z(x)=∑aρaeβ⟨x,za⟩.F(x)=\sum_a\rho_a e^{\beta\langle x,z_a\rangle}P_x^\perp(z_a),\ Z(x)=\sum_a\rho_a e^{\beta\langle x,z_a\rangle}. The normalized dynamics are x˙=F(x)/Z(x)\dot{x}=F(x)/Z(x). Because the representatives are fixed, Z(x)Z(x) is positive and bounded above and below on the sphere. Normalization therefore reparametrizes time along each token trajectory, preserving equilibria, attraction basins, and limiting clusters. Exponential convergence also persists, with adjusted rate bounds.
  • Implicit model. Under the same fixed-graph and cap assumptions, normalization preserves the interaction support and uniformly positive weights on existing edges. Cap invariance, componentwise consensus, and separation between components therefore persist, preserving the qualitative multiple-cluster conclusion. However, the convergence-rate bound requires a separate analysis because token-dependent normalization breaks the symmetry under fixed degree weights used in our proof.

We have added a discussion of these distinctions in updated submission. Please refer to Appendix A (Justification of Assumptions) for details.


[1] Chen, S., Lin, Z., Polyanskiy, Y., and Rigollet, P. Quantitative Clustering in Mean-Field Transformer Models. arXiv:2504.14697, 2025.

[2] Geshkovski, B., Letrouit, C., Polyanskiy, Y., and Rigollet, P. A Mathematical Perspective on Transformers. Bulletin (New Series) of the American Mathematical Society, 2025.

[3] Karagodin, N., Polyanskiy, Y., and Rigollet, P. Clustering in Causal Attention Masking. NeurIPS, 2024.

[4] Geshkovski, B., Letrouit, C., Polyanskiy, Y., and Rigollet, P. The emergence of clusters in self-attention dynamics. NeurIPS, 2023.

[5] Abella, Á. R., Silvestre, J. P., and Tabuada, P. Consensus is all you get: The role of attention in transformers. ICML, 2025.

[6] Bruno, G., Pasqualotto, F., and Agazzi, A. Emergence of Meta-Stable Clustering in Mean-Field Transformer Models. ICLR, 2025.

[7] Rigollet, P. The mean-field dynamics of transformers. ICM, 2026.

AuthorsOpenReview ↗

Response to Reviewer wVGc (Part 3)


3. Clustering scores in the original representation space (Q2)

We agree that evaluating clustering directly in the original representation space would strengthen the empirical evidence. However, a major practical challenge is the 16,384-dimensional representation space used in our DeepSeek-V4-Flash analysis. In very high-dimensional spaces, distance concentration can reduce the contrast between nearby and distant points, making local density estimation and density-based cluster identification more difficult. In our experiments, applying HDBSCAN directly to the original representations frequently produced fewer than two non-noise clusters, preventing meaningful computation of the corresponding silhouette scores. This outcome makes the direct HDBSCAN evaluation inconclusive under our current setup.

The original submission includes an analysis of the sensitivity of clustering scores to the reduced embedding dimension. Appendix G.4.1 (Table 8) reports results for 16, 32, 48, and 64 dimensions. For DeepSeek-V4-Flash on ARC-E, the scores ranges in 0.85–0.860.85–0.86, showing that the reported separation is stable and robust across the tested dimensions.


We thank the reviewer again for helping us clarify the scope of our guarantees and the interpretation of our empirical findings. Further discussions are more than welcome!

Comments on the paper

1 comment
AuthorsOpenReview ↗

To All Reviewers: Updated Manuscript and Appendix

We thank all reviewers for their constructive feedback. In addition to our individual responses, we have revised the manuscript and appendix to address the questions and suggestions raised. All revisions are marked in red. The main changes are summarized below:

  1. @Reviewer wVGc\textcolor{red}{\textbf{Reviewer wVGc}} and Reviewer bee6\textcolor{blue}{\textbf{Reviewer bee6}}: Justification of assumptions. Expanded the discussion of the motivations, roles, and limitations of our theoretical assumptions in Appendix A.

  2. @Reviewer bee6\textcolor{blue}{\textbf{Reviewer bee6}}: Trends and design implications. Added a discussion of how the framework helps characterize developments in sparse attention and identify complementary architectural design choices in Appendix A.

  3. @Reviewer bee6\textcolor{blue}{\textbf{Reviewer bee6}}: Categorization of modern LLMs. Added a main-text discussion categorizing the attention mechanisms used in modern frontier LLMs within our framework in Section 3.1.

  4. @Reviewer bee6\textcolor{blue}{\textbf{Reviewer bee6}}: Clustering across layers. Added an analysis of layerwise cosine silhouette scores in four LLMs in Appendix G.4.2.

  5. @Reviewer bee6\textcolor{blue}{\textbf{Reviewer bee6}} and Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}}: Cluster semantics. Added an analysis of the semantic and functional content of final-layer token clusters in Section 3.4.

  6. @Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}}: Comparison with prior theory. Clarified the differences in modeling assumptions and quantitative guarantees relative to existing clustering analyses in Appendix B.

  7. @Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}}: Fixed representatives and future work. Clarified the architectural motivation and analytical role of fixed representatives, and discussed extending the framework to evolving representatives in Appendix A, “Future Work.”

  8. @Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}}: Clustering before and after the final FFN. Added a comparison of clustering scores immediately before and after the final FFN in DeepSeek-V4-Flash in Appendix G.4.5.

  9. @Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}}: Basis of Corollary 2.1. Explicitly identified the Łojasiewicz convergence result underlying Corollary 2.1 in Section 2.1.

  10. @Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}}: Presentation of Assumption 2.2. Separated its concentration, separation, and representative-dominance conditions and clarified the associated notation in Section 2.1.

  11. @Reviewer hZkm\textcolor{green}{\textbf{Reviewer hZkm}}: Architectural categories in Figure 4. Revised the caption to explicitly state how each model’s attention mechanism is categorized within our framework.

We appreciate the reviewers’ suggestions, which have helped us clarify the contributions, practical relevance, and scope of the work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.