PaperlayerView exact revision on arXiv

Source-linked AI summary

Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation

Zunhai Su, Hengyuan Zhang, Wei Wu, Yifan Zhang, Yaxiu Liu, He Xiao, Qingyao Yang, Yuxuan Sun, Rui Yang, Chao Zhang, Jing Xiong, Hui Shen, Keyu Fan, Weihao Ye, Chaofan Tao, Taiqiang Wu, Zhongwei Wan, Tiantian Zhang, Bowen Yan, Zhen Li, Yiming Zhang, Congkai Xie, Yulei Qian, Yuchen Xie, Yik-Chung Wu, Hongxia Yang, Ngai Wong

arXiv:2604.10098v2cs.LG

TL;DR

Attention Sink concentrates attention on a few uninformative tokens, affecting Transformer dynamics and interpretability while aggravating hallucinations and robustness challenges. This survey synthesizes AS research across utilization, mechanistic interpretation, and mitigation. It concludes that AS can be leveraged, explained through several mechanisms, or controlled for robustness and efficiency, while important deployment and coverage limitations remain.

  • Problem

    Attention Sink is a persistent Transformer challenge, yet the growing literature lacked a comprehensive synthesis of its utilization, mechanisms, mitigation strategies, and research trajectory.

  • Method

    The survey systematically reviews AS studies across Transformer architectures and organizes them into Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation.

  • Results

    The review finds that AS influences training dynamics, interpretability, and inference efficiency, while utilization, mechanistic, and mitigation studies provide ways to leverage, explain, and control it.

  • Takeaways & Limitations

    AS research offers practical strategies for performance, robustness, and low-bit deployment, alongside mechanistic insights that motivate methods such as gated attention and modified softmax functions.

  • Takeaways & Limitations

    The survey primarily covers established architectures and does not comprehensively cover emerging or specialized models because AS studies for them remain limited.

Abstract

from arXiv · show

As the foundational architecture of modern machine learning, Transformers have driven remarkable progress across diverse AI domains. Despite their transformative impact, a persistent challenge across various Transformers is Attention Sink (AS), in which a disproportionate amount of attention is focused on a small subset of specific yet uninformative tokens. AS complicates interpretability, significantly affecting the training and inference dynamics, and exacerbates issues such as hallucinations. In recent years, substantial research has been dedicated to understanding and harnessing AS. However, a comprehensive survey that systematically consolidates AS-related research and offers guidance for future advancements remains lacking. To address this gap, we present the first survey on AS, structured around three key dimensions that define the current research landscape: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Our work makes a pivotal contribution by highlighting the key concepts and main trends in the field, guiding researchers through the evolution of AS-related studies. We envision this survey as a valuable resource, empowering researchers to effectively manage AS within the current Transformer paradigm, while simultaneously inspiring innovative advancements for the next generation of Transformers. The paper list of this work is available at https://github.com/ZunhaiSu/Awesome-Attention-Sink.

1. Introduction

Transformers have become foundational across AI, but Attention Sink concentrates attention on a few uninformative tokens, affecting dynamics, interpretability, hallucinations, and robustness. This survey organizes fragmented AS research into utilization, interpretation, and mitigation, tracing the field’s progression from empirical exploitation toward mechanistic understanding and strategic control.

  • Background: Transformers support progress across NLP, computer vision, multimodal learning, and embodied AI through long-range dependency modeling and scalable pretraining.
  • Background: Attention Sink concentrates disproportionate attention on a small set of uninformative tokens, influencing training and inference dynamics while complicating interpretability and exacerbating hallucinations and robustness challenges.
  • Research Questions: The literature addresses AS through three questions concerning fundamental utilization, mechanistic emergence and function, and architectures that operate independently of AS.
  • Position and Contributions: The survey responds to a fragmented literature lacking a definitive unified reference by systematically reviewing more than 200 studies and organizing them into Utilization, Interpretation, and Mitigation.
  • Survey Organization: Research has progressed from empirical AS utilization to mechanistic interpretation and, most recently, strategic mitigation, with the survey structuring these dimensions across its sections.
  • Resources: A maintained GitHub repository organizes the survey’s referenced papers and incorporates newer developments in AS research.

2. Attention Sink in Transformers

Transformers use self-attention as a general computational primitive, but across architectures they can develop Attention Sink: disproportionate attention to early or otherwise uninformative tokens. The survey characterizes its structural and training-related manifestations, including uses in MoE, vision-language-action, video diffusion, and 3D reconstruction models.

  • Architectural Overview: Transformer blocks combine multi-head self-attention with position-wise feed-forward networks, residual connections, and layer normalization.These components support long-range dependency modeling and stable training.
  • Cross-Architecture Manifestations: AS appears across autoregressive and classical language models, multimodal LLMs, MoE LLMs, and vision transformers despite architectural differences.In BERT, heads often focus on special tokens such as [CLS] and [SEP], while deeper heads can focus on punctuation.
  • Attention Sink Characterization: Attention Sink describes attention concentrating disproportionately on tokens with limited semantic or task-specific contribution.The defining combination is exceptionally high attention and intrinsically low-information content, rather than high attention alone.
  • Attention Sink Characterization: 98% of attention heads in LLaMA and other widely used LLMs frequently assign maximum attention to the first token, motivating threshold-based sink detection.A practical criterion classifies tokens whose cumulative attention substantially exceeds the global average as sinks.
  • Training Dynamics: AS emerges after sufficient optimization on adequate training data and becomes stronger with higher learning rates and substantial weight decay.Lower learning rates or minimal weight decay produce weaker or delayed sink formation, linking AS to training dynamics and head specialization.
  • Other Transformer Architectures: In MoE models, AS is tied to sparse Super Experts, while other applications preserve, repurpose, anchor, or suppress sinks for downstream objectives.Reported examples include 17.1% better real-world robotic manipulation, 6× video extrapolation, and stable kilometer-scale reconstruction at 18 FPS.

3. Fundamental Utilization of Attention Sink

The survey organizes Fundamental Utilization of Attention Sink into four paradigms that differ in how actively they manage or exploit sink tokens. These range from preserving naturally emerging sinks to redistributing attention, learning prefix tokens, and repurposing sinks.

  • Fundamental Utilization: Fundamental Utilization is organized into Sink Token Preservation, Attention Redistribution, Learnable Prefix Tokens, and Sink Token Repurposing.The section reviews each paradigm’s methodology, practical implementations, and key insights.
  • Strategy Spectrum: Sink Token Preservation passively maintains naturally emerging sink tokens without altering their attention distribution.Attention Redistribution instead actively reallocates attention from sink tokens toward semantically relevant regions.
  • Strategy Spectrum: Learnable Prefix Tokens use trainable tokens as a proactive strategy for managing or leveraging Attention Sink.The supplied passage introduces this paradigm but does not provide further implementation detail.
  • Strategy Spectrum: Sink Token Repurposing treats sink tokens as resources that can be redirected toward useful model functions.The supplied passage identifies this paradigm but does not specify its practical implementations.

3.1. Sink Token Preservation

Sink Token Preservation retains a small set of attention-absorbing tokens to stabilize attention under context compression. The strategy extends across KV caching, sparse attention, quantization, video, and multimodal systems, but dynamic sink positions remain difficult to detect efficiently.

  • Core Methodology: Sink Token Preservation permanently retains attention-absorbing tokens to stabilize attention during aggressive context compression.The retained sinks act as stable anchor points available to queries under causal attention.
  • KV Cache Compression: KV cache methods retain sink tokens while selectively evicting other tokens, including recent-window, heavy-hitter, hybrid, and adaptive strategies.StreamingLLM retains initial and recent tokens for stable infinite-length generation, while H2O preserves high-cumulative-attention heavy hitters.
  • Sparse Attention with Mask Enforcement: Sparse attention methods enforce sink-token visibility through masks or differentiated head-wise caching.MInference accelerates pre-filling by up to 10× without accuracy loss, while DuoAttention preserves full caches for retrieval heads and compresses streaming heads.
  • Quantization-Aware Protection: Pivot-token preservation protects sink tokens at full precision while quantizing other tokens, enabling 2-bit KV cache quantization with minimal performance loss.The protection targets sink tokens’ sensitivity to quantization-induced numerical degradation.
  • Cross-Modal and Video Extensions: Sink preservation generalizes across language, video, and multimodal models by retaining stable textual, frame, or visual attention anchors.Applications include multi-minute video generation without fine-tuning and preserving visual attention sinks during cross-modal fusion.
  • Discussion and Insights: Fixed sink positions are simple but may fail when sinks emerge elsewhere, whereas dynamic identification adds overhead and can conflict with optimized kernels.Efficient detection of non-initial sinks remains an open challenge across settings including ViTs and MLLMs.

3.2. Attention Redistribution

Attention Redistribution actively reduces attention assigned to sinks and reallocates the freed mass toward semantically relevant targets. Explicit rules offer predictable control, while adaptive calibration responds to input-specific sinks, with costs from identification and attention-score modification.

  • Core Methodology: Attention Redistribution reshapes attention by reducing sink influence and increasing focus on task-relevant tokens, unlike passive sink preservation.The paradigm includes explicit redistribution and attention-sink-aware calibration.
  • Explicit Redistribution: Explicit redistribution modifies sink and target scores using predefined parameters while preserving total per-query attention mass.The parameters determine retained sink mass and the proportion redirected to selected non-sink targets.
  • Explicit Redistribution: Full redistribution eliminates sink attention and redirects it to targets, supporting visual grounding, hallucination mitigation, and robustness against misleading inputs.Examples include VAR, AttnReal, GasEraser, positional-bias correction, and register-token approaches.
  • Explicit Redistribution: Sink reduction suppresses sink scores without explicit target redistribution, while head-level pattern broadcasting propagates visual anchoring across heads.VASparse reduces visual hallucinations while maintaining decoding efficiency, and EVAS enhances visual grounding while mitigating hallucinations.
  • Attention-Sink-Aware Calibration: Attention-sink-aware calibration dynamically evaluates sinks and adjusts attention distributions according to each input’s context.ACT adaptively selects harmful sinks and redistribution amounts, while scalar logit bias can indirectly control the full attention layout through Softmax normalization.
  • Discussion and Insights: Redistribution methods offer flexible control but require accurate sink or target identification and may incur latency, computational cost, and kernel-compatibility problems.Many methods modify post-Softmax scores, potentially limiting use with high-performance attention implementations.

3.3. Learnable Prefix Tokens

Learnable Prefix Tokens are trainable parameters inserted to act as explicit, stable attention sinks. They support streaming stability, vision-artifact mitigation, low-bit quantization, and cross-domain information aggregation, but require training and design tuning.

  • Core Methodology: Learnable Prefix Tokens are trainable parameters inserted into the input sequence to absorb excess attention as explicit sinks.Unlike natural sinks, they remain fixed during inference and provide predictable attention anchors.
  • Streaming Stability: A trainable placeholder retained in the KV cache provides stable attention over arbitrarily long streaming sequences.Unlike natural sinks, the placeholder cannot be evicted from the sliding window.
  • Quantization Facilitation: Learnable prefixes confine activation outliers to a small region, enabling per-tensor activation quantization without significant accuracy loss.CushionCache similarly uses tuned prefix tokens to mitigate later activation outliers under coarse-grained quantization.
  • Vision Artifact Mitigation: Register tokens absorb attention artifacts from low-information vision patches, producing cleaner attention maps across vision transformer models.The approach appears in pre-trained registers, post-hoc methods, and lightweight sink-token fine-tuning.
  • Information Aggregation: Learnable sink tokens can aggregate context for recommendation, long-context compression, code generation, and robotic spatial reasoning.RetoVLA reuses register outputs as dense global spatial context for action planning.
  • Discussion and Insights: Learnable Prefix Tokens provide predictable, controllable behavior across multiple domains but require additional training and empirical choices about number and placement.Their effectiveness depends on base-model capacity and training data, and cross-architecture generalization is not guaranteed.

3.4. Sink Token Repurposing

Sink Token Repurposing treats existing attention sinks as computational primitives rather than merely phenomena to preserve or suppress. The survey organizes applications into offensive, defensive, and efficiency-oriented paradigms spanning security, robustness, and optimization.

  • Core Methodology: Repurposing methods cover offensive, defensive, and efficiency-oriented applications across attack, defense, and optimization.The framework applies intrinsic AS properties without altering attention distributions or introducing additional tokens.
  • Core Methodology: Sink Token Repurposing exploits stable attention scores, fixed positions, numerical outliers, and geometric properties of existing AS.Unlike preservation, redistribution, or learnable prefix tokens, these methods use AS as computational primitives for specialized objectives.
  • Offensive Use: Attackers can inject triggers into AS positions or amplify AS attention to induce harmful behaviors.Backdoor unlearning places triggers at sink positions and aligns their attention values to enhance backdoor persistence.
  • Efficiency-Oriented Use: Efficiency-oriented methods identify AS as critical anchors or memory anchors to preserve important tokens and compress long contexts.KeyDiff preserves critical tokens while evicting redundant ones, OmniSparse prunes redundant video queries, and StreamingDialogue compresses conversation histories.
  • Discussion and Insights: AS provides a high-leverage intervention point for practical algorithms targeting security, robustness, and computational efficiency.The survey characterizes subtle AS manipulations as capable of producing substantial shifts in model behavior.
  • Limitations: Current methods often treat AS as static despite likely dynamic, context-dependent identities, magnitudes, and functional roles.The field also lacks rigorous theory for AS capacity, utility trade-offs, and downstream intervention impacts.

4. Mechanistic Interpretation of Attention Sink

The survey synthesizes mechanistic interpretations of AS across mathematical, numerical, functional, and geometric perspectives. These accounts connect Softmax normalization, negligible sink values, outlier circuits, and implicit attention biases to attention concentration and its mitigation.

  • Geometric Anchoring: The survey also treats geometric anchoring as an interpretation of AS’s influence within attention-space representation geometry.Together, the perspectives address mathematical origin, numerical mechanisms, functional role, and representational geometry.
  • Softmax Limitations and No-Op Theory: Softmax Limitations and No-Op Theory attributes AS to Softmax’s sum-to-one constraint when no key is meaningfully relevant.Attention concentrates on common low-information tokens whose negligible value vectors produce minimal residual updates.
  • Softmax Limitations and No-Op Theory: Sink tokens receive high attention but exhibit small value outputs, supporting the predicted no-op behavior across Transformer architectures.The survey cites evidence from BERT, ViT, LLaMA, and OLMo, including disproportionately high attention and near-zero value outputs.
  • Softmax Limitations and No-Op Theory: Relaxing or removing Softmax normalization mitigates or eliminates AS, providing causal support for the no-op theory.Gated attention reduces sink reliance, while independently applied sigmoid attention removes forced attention allocation.

4.2. Outlier Circuits

The Outlier Circuits perspective explains AS through systematic weight, activation, and attention outliers that form interconnected pathways across layers. Evidence links these outliers causally to AS, while the framework’s training dynamics and component interactions remain incompletely understood.

  • Core Concepts: Outlier Circuits identify systematic outliers as numerical infrastructure sustaining AS through circuit-like pathways.The framework complements Softmax-based explanations by addressing the mechanisms that maintain attention sinks.
  • Types of Systematic Outliers: Weight, activation, and attention outliers are interdependent: weight outliers align with activation outliers, which coincide with AS across sequence positions.Attention outliers correspond to keys receiving disproportionately high cumulative attention and persist across nearly all layers.
  • Formation and Evolution: Activation outliers increase query-key alignment and dot products, causing Softmax to assign disproportionately high attention to sink tokens.This forms circuit-like pathways that stabilize AS across the model’s layers.
  • Formation and Evolution: AS follows a predictable cross-layer lifecycle, emerging in early layers, stabilizing in middle layers, and gradually vanishing in final layers.This lifecycle is documented for extreme activation outliers in LLaMA2-7B.
  • Causal Evidence: Pruning only three of 6,144 Super Experts causes catastrophic AS collapse and repetitive, uninformative outputs.The intervention provides causal evidence that outlier sources are functionally central to AS behavior.
  • Limitations: The framework’s causal validation remains incomplete, and training dynamics governing circuit emergence, stabilization, and evolution are largely unexplored.The survey calls for systematic interventions and longitudinal analyses across training epochs.

4.3. Implicit Attention Bias

Implicit Attention Bias interprets AS as a stable bias added to attention outputs, explaining high sink attention despite minimal output contribution. Explicit key or value biases and context-aware scaling can replace or eliminate this implicit mechanism, although its training dynamics remain unresolved.

  • Core Concepts: Implicit Attention Bias treats AS as a fixed, input-independent bias term within the attention output.This perspective focuses on AS’s functional role rather than its mathematical origin or numerical implementation.
  • Core Concepts: AS-token value updates remain nearly identical across query positions and inputs, functioning as a constant bias added to every attention output.The token set C contains Massive Activation tokens identified with AS.
  • Causal Evidence: Training GPT-2 with explicit learnable key and value biases makes Massive Activations disappear and eliminates AS.This supports the interpretation that AS is an implicit bias learned to cope with the Softmax constraint.
  • Observational Evidence: Across every attention head, aggregate value updates remain highly consistent across tokens whenever attention sinks emerge.KVSink reports this consistency as evidence that the activations act as stable, input-independent attention biases.
  • Causal Evidence: Learnable key biases can shift the sink position, while context-aware scaling prevents systematic outliers and eliminates AS.These interventions provide complementary causal evidence that AS functions as an implicit attention bias.
  • Limitations: The training dynamics producing Massive Activations and AS remain unformalized, and relationships among identified implicit biases are fragmented.It is also unknown whether more general or effective forms of implicit bias exist.

4.4. Geometric Anchoring

The Geometric Anchoring perspective treats attention sinks as stable reference points that organize token representations and attention in high-dimensional spaces. Evidence supports practical uses in context extension, KV-cache efficiency, and selective attention, but mechanistic causality remains insufficiently established.

  • Core Concept: Geometric Anchoring models sink tokens as stable reference points that structure the representational geometry of other tokens.The perspective interprets attention sinks as a geometric phenomenon rather than solely as Softmax artifacts or activation outliers.
  • Supporting Evidence: After the first layer, initial tokens exhibit distinct positional vectors while later tokens converge toward similar representations.This positional-vector pattern supports the sink token’s anchoring role.
  • Additional Applications: Code-generation attention can concentrate on structural anchors, with the top two attention weights often exceeding 80%.Other approaches use dedicated or structural sink tokens as position-independent or model-specific anchors.
  • Supporting Evidence: Sink tokens remain geometrically stable across layers, while other tokens increasingly resemble the sink after attention-sink emergence.The reported cosine-similarity patterns characterize the sink token as static relative to other token representations.
  • Advantages: Geometric properties support positional-vector replacement for context extension, orthogonality-based pruning, and key-similarity-based KV-cache eviction.The survey reports that key-similarity-based eviction often outperforms conventional attention-score-based methods.
  • Limitations: The framework’s evidence is mostly correlational, leaving key mechanistic claims unvalidated and the emergence of particular anchors unexplained.It also does not fully explain how anchors interact with broader model dynamics.

4.5. Other Mechanistic Interpretations

Other interpretations explain attention sinks through mathematical constraints, architectural bias, training dynamics, numerical mechanisms, geometric structure, and functional roles. Together, these perspectives frame attention sinks as arising from interacting causes that affect information mixing, representation stability, and attention behavior.

  • Geometric Structure: Anti-Overmixing Theory casts the first token as a residual-stream anchor that limits excessive information mixing and prevents representational collapse.The associated account links the sink to preserving contextual distinctiveness across layers.
  • Geometric Structure: Spectral-Energy Association Theory describes the first token as a large-norm dark signal that dominates attention energy and compresses the representational manifold.The theory attributes attention sinks to low-rank spectral dynamics that trade token distinctions for stable information processing.
  • Synthesis: The survey consolidates these interpretations across five analytical levels to explain attention-sink emergence, persistence, and impact.The levels span mathematical origin, training dynamics, numerical mechanism, geometric structure, and functional role.
  • Mathematical Origin: Softmax’s sum-to-one constraint can force attention onto uninformative tokens when no meaningful key is available.The mechanism lacks a natural null option for queries without semantically relevant keys.
  • Structural Bias: Causal masking gives early tokens cumulative visibility, while RoPE’s distance-dependent decay can create biases toward sink tokens.The survey identifies both mechanisms as architectural sources of attention-sink formation.
  • Training Dynamics: Active-Dormant Attention Theory attributes sink specialization to heads with large key norms and small value norms reinforced by positive gradient feedback.This account focuses on how attention sinks emerge and stabilize during training.
  • Numerical Mechanism: Outlier Circuits and Mix-Compress-Refine theories describe numerical pathways that stabilize sinks and compress contextual information before selective refinement.The survey also places outlier-driven rescaling among the numerical mechanisms associated with attention sinks.
  • Functional Role: Implicit Attention Bias interprets attention sinks as fixed, input-independent bias terms because sink-token value updates are nearly identical across queries and inputs.This interpretation assigns a functional role to the sink beyond token-specific semantic content.

5. Strategic Mitigation of Attention Sink

Strategic mitigation approaches address Attention Sink by decoupling no-op behavior from extreme Softmax logits, modifying normalization, or introducing explicit attention biases. These methods can reduce AS and activation outliers, but several require training from scratch and involve trade-offs in parameters, dynamics, or evaluation.

  • Gated Attention Mechanisms: Gated Attention introduces learnable gates that directly suppress attention outputs, breaking the self-reinforcing cycle that produces Attention Sink.Output gating after SDPA and value-state gating are the two primary strategies.
  • Limitations: Gated Attention, modified Softmax, and learnable-bias methods generally require training from scratch, while their other limitations include parameter overhead, excessive flattening, and nonstandardized evaluation.The approaches also leave training dynamics and interactions with attention distributions incompletely understood.
  • Modified Softmax Functions: Modified Softmax Functions intervene directly in normalization to prevent extreme logits and forced attention allocation without adding parameters.Output-constrained Softmax, normalization-free attention, and pre-Softmax modulation are the main families described.
  • Modified Softmax Functions: Several modified Softmax variants achieve strong AS mitigation: Softmax-1 reduces first-token attention from 65% to 3.3%, while Softpick reports a 0% sink rate.Softmax-1 also lowers activation kurtosis from 1657 to 3.1; Softpick lowers it from 33,510 to 340.
  • Learnable Attention Bias: Learnable Attention Bias replaces AS-induced implicit bias with trainable explicit biases that control no-update attention and can eliminate sink behavior.Key biases make AS disappear and attach it to the bias position, supporting the interpretation that AS can substitute for an explicit key bias.

6. Applications and Practical Guidelines

The survey organizes practical AS guidance by application domain, matching techniques to architectures and task requirements. Recommended strategies either preserve or repurpose beneficial sinks, redistribute harmful attention, or modify attention mechanisms to mitigate AS.

  • Applications and Practical Guidelines: Application-specific guidance selects AS techniques according to model architecture and task requirements.The survey categorizes research by application domain and provides concrete recommendations for managing AS.
  • Learnable Prefix Tokens: Learnable Prefix Tokens absorb sink artifacts, stabilize ViT optimization, support low-bit quantization, and create dedicated anchors for training-based extensions.They are also recommended for domain-specific tasks and can be injected at test time in some ViT-based settings.
  • Long-Context Enhancement: Sink Token Preservation retains initial AS tokens as fixed anchors for KV-cache compression, sparse attention, and long-context streaming.These anchors enable aggressive eviction or quantization without performance collapse and support stable sparse-attention patterns.
  • Multimodal and Training-Free Applications: Attention Redistribution shifts attention from visual, OCR, or vocabulary sink tokens toward informative regions or tokens to address hallucination and factual-generation issues.The survey also describes redistribution for training-free improvement and position-bias mitigation.
  • Strategic Mitigation and Repurposing: Gated Attention Mechanisms and Modified Softmax Functions can remove or prevent AS, while Sink Token Repurposing preserves useful patterns or detects model corruption.Repurposing is also described for dialogue AS, spatial memory, backdoor gateways, and fine-tuning-related corruption indicators.

7. Challenges and Future Directions

The survey identifies unresolved challenges in computational efficiency, adaptation cost, training-dynamics theory, architectural scope, evaluation, transfer, and combining techniques. It proposes lightweight adaptation, formalized theory, standardized benchmarks, cross-architecture studies, and synergistic methods as future directions.

  • Challenges and Future Directions: Dynamic sink detection and Attention Redistribution add computational overhead, while post-Softmax operations can limit compatibility with high-performance attention kernels.These costs may become bottlenecks in large-scale models.
  • Challenges and Future Directions: Most Gated Attention, Modified Softmax, Learnable Attention Bias, and Learnable Prefix Token methods require training from scratch or additional fine-tuning.Their adaptation cost limits practical adoption for already pretrained large models.
  • Challenges and Future Directions: Softmax Limitations and No-Op Theory does not fully explain the training dynamics that produce no-op behavior, mutual reinforcement, and systematic outlier alignment.The interactions among attention scores, value states, weights, activations, and attention outliers remain incompletely formalized.
  • Lightweight Adaptation for Pre-trained Models: Future work should develop parameter-efficient AS adaptation, including adapters, LoRA, and continual pre-training with inserted AS-aware components.These approaches aim to avoid full retraining of pretrained models.
  • Theoretical Formalization: A unified theoretical framework could integrate Softmax constraints, optimization dynamics, implicit bias, and outlier-circuit perspectives to guide principled mitigation.The survey also calls for formalizing AS emergence, evolution, and functional roles.
  • Evaluation and Generalization: Standardized benchmarks are needed to compare AS mitigation methods fairly across efficacy, computational overhead, parameter introduction, and related factors.The survey further identifies cross-architecture transfer, emerging architectures, and coordinated technique combinations as open research directions.

8. Conclusion

The survey synthesizes AS research across utilization, interpretation, and mitigation, showing its relevance to training dynamics, interpretability, inference efficiency, robustness, and low-bit deployment. It also identifies persistent practical and theoretical challenges and outlines future research priorities.

  • Conclusion: The survey synthesizes over 180 studies across Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation.It presents the first comprehensive survey of AS in Transformer architectures.
  • Conclusion: AS profoundly influences training dynamics, model interpretability, and inference efficiency across diverse Transformer architectures.The survey connects empirical utilization, mechanistic explanation, and mitigation to these effects.
  • Conclusion: Utilization strategies leverage AS, mechanistic studies examine its causes and roles, and mitigation techniques control or suppress it for robustness and low-bit deployment.These three dimensions form the survey’s integrated account of AS research.
  • Conclusion: Remaining challenges include computational efficiency, training-from-scratch requirements, and incomplete understanding of training dynamics.The survey highlights efficient lightweight handling, emerging architectures, and standardized benchmarks as promising directions.

9. Limitations

The survey primarily covers well-established Transformer architectures with substantial prior AS research. Emerging or specialized architectures are not comprehensively covered because relevant studies remain limited.

  • Limitations: The analysis focuses primarily on CLMs, LLMs, MLLMs, MoE LLMs, and ViTs.These architectures have been extensively studied in prior literature.
  • Limitations: Hybrid-linear attention models, VGGT, and other emerging or specialized architectures are not comprehensively covered.The stated reason is the limited availability of AS-related studies for these architectures.
  • Limitations: The survey plans to incorporate relevant studies as research on novel architectures expands.The authors state that the presented insights and methodologies may inform understanding across other architectures.

A. Comprehensive Overview of Surveyed Papers

Table 1 surveys papers across three Attention Sink dimensions: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Most studies address only a subset of these dimensions, marked by “-” where a dimension is absent.

  • Table 1 annotates surveyed papers by Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation of AS.
  • The table’s central comparison is which AS dimensions each surveyed paper addresses.
  • A “-” indicates that a surveyed paper does not target the corresponding dimension.
Loading 2604.10098v2…