What are you sinking? A geometric approach on attention sink

• URL: http://arxiv.org/abs/2508.02546v1
• Date: Mon, 04 Aug 2025 15:59:15 GMT
• Title: What are you sinking? A geometric approach on attention sink
• Abstract summary: Attention sink (AS) is a consistent pattern in transformer attention maps where certain tokens disproportionately attract attention from other tokens.<n>We show that in transformers, AS is not an architectural artifact, but it is the manifestation of a fundamental geometric principle.
• Score: 6.552700667389349
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Attention sink (AS) is a consistent pattern in transformer attention maps where certain tokens (often special tokens or positional anchors) disproportionately attract attention from other tokens. We show that in transformers, AS is not an architectural artifact, but it is the manifestation of a fundamental geometric principle: the establishment of reference frames that anchor representational spaces. We analyze several architectures and identify three distinct reference frame types, centralized, distributed, and bidirectional, that correlate with the attention sink phenomenon. We show that they emerge during the earliest stages of training as optimal solutions to the problem of establishing stable coordinate systems in high-dimensional spaces. We show the influence of architecture components, particularly position encoding implementations, on the specific type of reference frame. This perspective transforms our understanding of transformer attention mechanisms and provides insights for both architecture design and the relationship with AS.

Related papers

• Global Divergence, Local Convergence: Representation Geometry in SSMs and Transformers [4.71326501896672]
  Recent state-space models (SSMs) such as Mamba achieve language modeling performance comparable to transformers.<n>This raises an important question: how do these structural differences influence the geometry and functional nature of their internal representations?<n>We study this question through a multi-scale analysis of representations in transformers, SSMs, and hybrid architecture.
  arXiv  Detail & Related papers  (2026-09-08T12:57:30Z)
• Beyond Point-Wise Matching: Structural Representation Alignment for Accelerating Diffusion Transformers [93.3976834364707]
  We propose sREPA, a structural REPresentation Alignment framework to enforce consistency in the relational geometry of feature maps.<n>By encouraging the model to internalize holistic spatial layouts and structural correlations from pre-trained features, sREPA achieves faster and more stable convergence.
  arXiv  Detail & Related papers  (2026-05-16T12:01:04Z)
• A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation [1.2744523252873352]
  Transformers commonly exhibit an attention sink: disproportionately high attention to the first position.<n>We study this behavior in GPT-2-style models with learned query biases and absolute positional embeddings.
  arXiv  Detail & Related papers  (2026-04-16T07:32:37Z)
• ASTRA: Enhancing Multi-Subject Generation with Retrieval-Augmented Pose Guidance and Disentangled Position Embedding [1.4145309506833208]
  ASTRA is a novel framework that architecturally disentangles subject appearance from pose structure.<n>ASTRA achieves a new state-of-the-art in pose adherence, while maintaining high identity fidelity and text alignment in DreamBench.
  arXiv  Detail & Related papers  (2026-04-15T14:50:05Z)
• Learning 3D Representations for Spatial Intelligence from Unposed Multi-View Images [81.94999489820974]
  UniSplat is a feed-forward framework for learning 3D representations from unposed multi-view images.<n>We introduce a dual-masking strategy that strengthens geometry induction in the encoder.<n>Second, we develop a coarse-to-fine Gaussian splatting strategy that reduces appearance-semantics inconsistencies.<n>Third, we introduce a pose-conditioned recalibration mechanism that interrelates predicted 3D point and semantic maps into the image plane.
  arXiv  Detail & Related papers  (2026-04-12T10:36:18Z)
• How Attention Sinks Emerge in Large Language Models: An Interpretability Perspective [67.08306259095778]
  We identify a simple mechanism that enables the model to recognize token at position zero and induce an attention sink within two transformer blocks.<n>By analyzing training traces from a 30B A3B MoE model trained from scratch, we find that this mechanism emerges early in training and becomes increasingly concentrated in the first two layers.
  arXiv  Detail & Related papers  (2026-02-04T11:10:34Z)
• Attention as Binding: A Vector-Symbolic Perspective on Transformer Reasoning [0.0]
  Transformer-based language models display impressive reasoning-like behavior, yet remain brittle on tasks that require stable symbolic manipulation.<n>This paper develops a unified perspective on these phenomena by interpreting self-attention and residual streams as implementing an approximate Vector Symbolic Architecture (VSA)<n>In this view, queries and keys define role spaces, values encode fillers, attention weights perform soft unbinding, and residual connections realize superposition of many bound structures.
  arXiv  Detail & Related papers  (2025-12-08T05:38:24Z)
• GeoSketch: A Neural-Symbolic Approach to Geometric Multimodal Reasoning with Auxiliary Line Construction and Affine Transformation [28.500787311066563]
  GeoSketch is a neural-symbolic framework that recasts geometric reasoning as an interactive perception-reasoning-action loop.<n>By unifying hierarchical decision-making, executable visual actions, and symbolic verification, GeoSketch advances multimodal reasoning from static interpretation to dynamic interaction.
  arXiv  Detail & Related papers  (2025-09-26T15:12:04Z)
• Cross-architecture universal feature coding via distribution alignment [88.73189953617594]
  We introduce a new research problem: cross-architecture universal feature coding (CAUFC)<n>We propose a two-step distribution alignment method. First, we design the format alignment method that CNN and Transformer features into a consistent 2D token format. Second, we propose the feature value alignment method that harmonizes statistical distributions via truncation and normalization.<n>As a first attempt to study CAUFC, we evaluate our method on the image classification task. Experimental results demonstrate that our method achieves superior rate-accuracy trade-offs compared to the architecture-specific baseline.
  arXiv  Detail & Related papers  (2025-06-15T06:14:02Z)
• On the Emergence of Position Bias in Transformers [59.87743433861665]
  This paper presents a graph-theoretic framework for analyzing position biases in multilayer positions.<n>Our framework offers a principled foundation for understanding positional interplay in transformers.
  arXiv  Detail & Related papers  (2025-02-04T02:53:07Z)
• Learning Correlation Structures for Vision Transformers [93.22434535223587]
  We introduce a new attention mechanism, dubbed structural self-attention (StructSA) We generate attention maps by recognizing space-time structures of key-query correlations via convolution. This effectively leverages rich structural patterns in images and videos such as scene layouts, object motion, and inter-object relations.
  arXiv  Detail & Related papers  (2024-04-05T07:13:28Z)
• GTA: A Geometry-Aware Attention Mechanism for Multi-View Transformers [63.41460219156508]
  We argue that existing positional encoding schemes are suboptimal for 3D vision tasks. We propose a geometry-aware attention mechanism that encodes the geometric structure of tokens as relative transformation. We show that our attention, called Geometric Transform Attention (GTA), improves learning efficiency and performance of state-of-the-art transformer-based NVS models.
  arXiv  Detail & Related papers  (2023-10-16T13:16:09Z)
• Spherical Position Encoding for Transformers [0.0]
  We introduce the notion of "geotokens" which are input elements for transformer architectures. Unlike the natural language the sequential position is not important for the model but the geographical coordinates are. We formulate a position encoding mechanism based on RoPE architecture which is adjusted for spherical coordinates.
  arXiv  Detail & Related papers  (2023-10-04T09:28:59Z)
• On the interplay of adversarial robustness and architecture components: patches, convolution and attention [65.20660287833537]
  We study the effect of adversarial training on the interpretability of the learnt features and robustness to unseen threat models. An ablation from ResNet to ConvNeXt reveals key architectural changes leading to almost $10%$ higher $ell_infty$-robustness.
  arXiv  Detail & Related papers  (2022-09-14T22:02:32Z)
• Ripple Attention for Visual Perception with Sub-quadratic Complexity [7.425337104538644]
  Transformer architectures are now central to modeling in natural language processing tasks. We propose ripple attention, a sub-quadratic attention mechanism for visual perception. In ripple attention, contributions of different tokens to a query are weighted with respect to their relative spatial distances in the 2D space.
  arXiv  Detail & Related papers  (2021-10-06T02:00:38Z)
• Cross-view Geo-localization with Evolving Transformer [7.5800316275498645]
  Cross-view geo-localization is challenging due to drastic appearance and geometry differences across views. We devise a novel geo-localization Transformer (EgoTR) that utilizes the properties of self-attention in Transformer to model global dependencies. Our EgoTR performs favorably against state-of-the-art methods on standard, fine-grained and cross-dataset cross-view geo-localization tasks.
  arXiv  Detail & Related papers  (2021-07-02T05:33:14Z)
• Twins: Revisiting Spatial Attention Design in Vision Transformers [81.02454258677714]
  In this work, we demonstrate that a carefully-devised yet simple spatial attention mechanism performs favourably against the state-of-the-art schemes. We propose two vision transformer architectures, namely, Twins-PCPVT and Twins-SVT. Our proposed architectures are highly-efficient and easy to implement, only involving matrix multiplications that are highly optimized in modern deep learning frameworks.
  arXiv  Detail & Related papers  (2021-04-28T15:42:31Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.