mlsanityData to 7 Oct 2026

Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration

arXiv 2406.15765 ↗First posted 22 Jun 2024cs.LG cs.CLcited by 70

Not on the map. Why?

Abstract (arXiv)

Attention is a fundamental component behind the remarkable achievements of large language models (LLMs). However, our current understanding of the attention mechanism, especially regarding how attention distributions are established, remains limited. Inspired by recent studies that explore the presence of attention sink in the initial token, which receives disproportionately large attention scores despite their lack of semantic importance, this work delves deeper into this phenomenon. We aim to provide a more profound understanding of the existence of attention sinks within LLMs and to uncover ways to enhance the achievable accuracy of LLMs by directly optimizing the attention distributions, without the need for weight finetuning. Specifically, this work begins with comprehensive visualizations of the attention distributions in LLMs during inference across various inputs and tasks. Based on these visualizations, to the best of our knowledge, we are the first to discover that (1) attention sinks occur not only at the start of sequences but also within later tokens of the input, and (2) not all attention sinks have a positive impact on the achievable accuracy of LLMs. Building upon our findings, we propose a training-free Attention Calibration Technique (ACT) that automatically optimizes the attention distributions on the fly during inference in an input-adaptive manner. Extensive experiments validate that ACT consistently enhances the accuracy of various LLMs across different applications. Specifically, ACT achieves an average improvement of up to 7.30% in accuracy across different datasets when applied to Llama-30B. Our code is available at https://github.com/GATECH-EIC/ACT.

Key facts. Cited by 70 papers in the map (data through 7 Oct 2026). Builds on 44 of its 59 references that are in the map. First posted on arXiv 22 Jun 2024.

Builds on

44 of 59 references.

References · 44

  1. 2017Attention Is All You Needcited by 59,932
  2. 2020Language Models are Few-Shot Learnerscited by 32,199
  3. 2023GPT-4 Technical Reportcited by 18,132
  4. 2022Training language models to follow instructions with human feedbackcited by 15,712
  5. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  6. 2023LLaMA: Open and Efficient Foundation Language Modelscited by 13,958
  7. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  8. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  9. 2020Measuring Massive Multitask Language Understandingcited by 7,768
  10. 2020Scaling Laws for Neural Language Modelscited by 6,411
  11. 2018Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challengecited by 4,710
  12. 2016SQuAD: 100,000+ Questions for Machine Comprehension of Textcited by 4,115
32 more
  1. 2019HellaSwag: Can a Machine Really Finish Your Sentence?cited by 4,018
  2. 2021The Power of Scale for Parameter-Efficient Prompt Tuningcited by 3,769
  3. 2022OPT: Open Pre-trained Transformer Language Modelscited by 3,464
  4. 2023QLoRA: Efficient Finetuning of Quantized LLMscited by 3,355
  5. 2023Mistral 7Bcited by 3,065
  6. 2019PIQA: Reasoning about Physical Commonsense in Natural Languagecited by 2,818
  7. 2015Character-level Convolutional Networks for Text Classificationcited by 2,293
  8. 2019BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questionscited by 2,140
  9. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,115
  10. 2018Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answeringcited by 2,114
  11. 2022BLOOM: A 176B-Parameter Open-Access Multilingual Language Modelcited by 1,968
  12. 2018CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledgecited by 1,793
  13. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,734
  14. 2019SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systemscited by 1,718
  15. 2018Know What You Don't Know: Unanswerable Questions for SQuADcited by 1,538
  16. 2021Multitask Prompted Training Enables Zero-Shot Task Generalizationcited by 1,462
  17. 2021GLM: General Language Model Pretraining with Autoregressive Blank Infillingcited by 1,026
  18. 2019What Does BERT Look At? An Analysis of BERT's Attentioncited by 994
  19. 2019A Multiscale Visualization of Attention in the Transformer Modelcited by 292
  20. 2022LST: Ladder Side-Tuning for Parameter and Memory Efficient Transfer Learningcited by 228
  21. 2023LoftQ: LoRA-Fine-Tuning-Aware Quantization for Large Language Modelscited by 195
  22. 2022Scaling Up Models and Data with textttt5x and textttseqiocited by 187
  23. 2023LoRA-FA: Efficient and Effective Low Rank Representation Fine-tuningcited by 132
  24. 2023Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMscited by 82
  25. 2022Structured Prompting: Scaling In-Context Learning to 1,000 Examplescited by 81
  26. 2024Chain of LoRA: Efficient Fine-tuning of Language Models via Residual Learningcited by 80
  27. 2023GPT4AIGChip: Towards Next-Generation AI Accelerator Design Automation via Large Language Modelscited by 59
  28. 2024APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inferencecited by 29
  29. 2023Master-ASR: Achieving Multilingual Scalability and Low-Resource Adaptation in ASR with Modular Learningcited by 9
  30. 2022Losses Can Be Blessings: Routing Self-Supervised Speech Representations Towards Efficient Multilingual and Multitask Speech Processingcited by 7
  31. 2023PILLOW: Enhancing Efficient Instruction Fine-tuning via Prompt Matchingcited by 6
  32. 2023Hint-Aug: Drawing Hints from Foundation Vision Transformers Towards Boosted Few-Shot Parameter-Efficient Tuningcited by 2

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 9; 2025: 32; 2026: 29.

70 papers.

Sort:

By area:

  1. 2025Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Freecited by 266
  2. 2024When Attention Sink Emerges in Language Models: An Empirical Viewcited by 200
  3. 2025See What You Are Told: Visual Attention Sink in Large Multimodal Modelscited by 127
  4. 2024Model Tells You Where to Merge: Adaptive KV Cache Merging for LLMs on Long-Context Taskscited by 94
  5. 2025Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areascited by 93
  6. 2024Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMscited by 54
  7. 2024Seeing Clearly by Layer Two: Enhancing Attention Heads to Alleviate Hallucination in LVLMscited by 49
  8. 2025Seed-Coder: Let the Code Model Curate Data for Itselfcited by 46
  9. 2025RotateKV: Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotationscited by 42
  10. 2025What are you sinking? A geometric approach on attention sinkcited by 26
  11. 2026The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinkscited by 23
  12. 2026A Unified View of Attention and Residual Sinks: Outlier-Driven Rescaling is Essential for Transformer Trainingcited by 21
38 more
  1. 2025To Sink or Not to Sink: Visual Information Pathways in Large Vision-Language Modelscited by 18
  2. 2025Attention Sinks: A 'Catch, Tag, Release' Mechanism for Embeddingscited by 16
  3. 2025Mirage in the Eyes: Hallucination Attack on Multi-modal Large Language Models with Only Attention Sinkcited by 15
  4. 2025Interpreting the Repeated Token Phenomenon in Large Language Modelscited by 15
  5. 2025ParallelComp: Parallel Long-Context Compressor for Length Extrapolationcited by 13
  6. 2025ZeroTuning: Unlocking the Initial Token's Power to Enhance Large Language Models Without Trainingcited by 11
  7. 2026Attention Sinks Are Provably Necessary in Softmax Transformers: Evidence from Trigger-Conditional Taskscited by 10
  8. 2024House of Cards: Massive Weights in LLMscited by 9
  9. 2025Improving Complex Reasoning with Dynamic Prompt Corruption: A soft prompt Optimization Approachcited by 9
  10. 2025How Do Large Vision-Language Models See Text in Image? Unveiling the Distinctive Role of OCR Headscited by 9
  11. 2025Mitigating Attention Sinks and Massive Activations in Audio-Visual Speech Recognition with LLMscited by 9
  12. 2025Identifying and Evaluating Inactive Heads in Pretrained LLMscited by 8
  13. 2025OrthoRank: Token Selection via Sink Token Orthogonality for Efficient LLM inferencecited by 8
  14. 2025Lost in the Middle: An Emergent Property from Information Retrieval Demands in LLMscited by 8
  15. 2025All You Need is One: Capsule Prompt Tuning with a Single Vectorcited by 8
  16. 2026Hallucination Begins Where Saliency Dropscited by 8
  17. 2025Attention Mechanisms Perspective: Exploring LLM Processing of Graph-Structured Datacited by 7
  18. 2025Value-State Gated Attention for Mitigating Extreme-Token Phenomena in Transformerscited by 7
  19. 2025Don't Deceive Me: Mitigating Gaslighting through Attention Reallocation in LMMscited by 6
  20. 2026Context Tokens are Anchors: Understanding the Repetition Curse in dMLLMs from an Information Flow Perspectivecited by 6
  21. 2026On the Nature of Attention Sink that Shapes Decoding Strategy in Omni-LLMscited by 5
  22. 2024EMS: Adaptive Evict-then-Merge Strategy for Head-wise KV Cache Compression Based on Global-Local Importancecited by 4
  23. 2026Why Attention Patterns Exist: A Unifying Temporal Perspective Analysiscited by 4
  24. 2026Attention Sink Forges Native MoE in Attention Layers: Sink-Aware Training to Address Head Collapsecited by 4
  25. 2024AERO: Entropy-Guided Framework for Private LLM Inferencecited by 3
  26. 2025Gradient-guided Attention Map Editing: Towards Efficient Contextual Hallucination Mitigationcited by 3
  27. 2025Enhancing Chain-of-Thought Reasoning with Critical Representation Fine-tuningcited by 3
  28. 2025CTR-Sink: Attention Sink for Language Models in Click-Through Rate Predictioncited by 3
  29. 2025Integral Transformer: Denoising Attention, Not Too Much Not Too Littlecited by 3
  30. 2025On the Existence and Behavior of Secondary Attention Sinkscited by 3
  31. 2026Surgery: Mitigating Harmful Fine-Tuning for Large Language Models via Attention Sinkcited by 3
  32. 2026OccamToken: Efficient VLM Inference with Training-Free and Budget-Adaptive Token Pruningcited by 3
  33. 2025ReasoningGuard: Safeguarding Large Reasoning Models with Inference-time Safety Aha Momentscited by 2
  34. 2026What Makes Position Zero Special? A Mechanistic Study of Position Zero Attention Sinks in LLMscited by 2
  35. 2026When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Modelscited by 2
  36. 2026Massive Spikes in LLMs are Bias Vectors: Mechanistic Uncovering and Spike-Free Quantizationcited by 2
  37. 2024Enhanced Structured State Space Models via Grouped FIR Filtering and Attention Sink Mechanismscited by 1
  38. 2025Probability Consistency in Large Language Models: Theoretical Foundations Meet Empirical Discrepanciescited by 1

50 most cited of 70 All 70 (JSON)