Pith. sign in

REVIEW 9 cited by

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

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2406.15765 v1 pith:3C6KV2MN submitted 2024-06-22 cs.LG cs.CL

Zhongzhi Yu , Zheng Wang , Yonggan Fu , Huihong Shi , Khalid Shaikh , Yingyan Celine Lin This is my paper · ORCID
classification cs.LGcs.CL
keywords attentionllmsaccuracydistributionssinksacrosslargeachievable
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
read the original abstract

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.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models

    cs.CV 2026-04 unverdicted novelty 7.0 of 10

    Layer-wise Sink Gating scales vision and LLM attention sinks in LVLMs to balance global priors and local evidence, improving multimodal benchmarks with a frozen backbone.

  2. Attention Mechanisms Perspective: Exploring LLM Processing of Graph-Structured Data

    cs.AI 2025-05 conditional novelty 7.0 of 10

    Fine-tuning LLMs on graph tasks with a restricted attention window that includes partial graph links, then switching to a fully connected window at inference, improves node classification over standard full-attention ...

  3. Mirage in the Eyes: Hallucination Attack on Multi-modal Large Language Models with Only Attention Sink

    cs.LG 2025-01 conditional novelty 7.0 of 10

    Adversarial images optimized to induce attention sink behavior increase hallucination rates in multiple MLLMs, including commercial APIs, without visibly degrading response quality.

  4. What Makes Position Zero Special? A Mechanistic Study of Position Zero Attention Sinks in LLMs

    cs.LG 2026-02 conditional novelty 6.0 of 10

    Position-zero attention sinks in transformers emerge from causal-masking asymmetry: position zero attends only to itself, and an MLP then amplifies its representation into a stable, high-norm 'sink'.

  5. RotateKV: Accurate and Robust 2-Bit KV Cache Quantization for LLMs via Outlier-Aware Adaptive Rotations

    cs.LG 2025-01 conditional novelty 6.0 of 10

    RotateKV combines outlier-aware, pre-RoPE grouped-head Hadamard rotation with attention-sink-aware retention to make 2-bit KV cache quantization accurate on LLaMA-2, LLaMA-3, and Mistral models.

  6. Enhancing Chain-of-Thought Reasoning with Critical Representation Fine-tuning

    cs.CL 2025-07 conditional novelty 5.0 of 10

    CRFT selects critical internal representations via attention and saliency scores and fine-tunes only them, improving GSM8K accuracy over ReFT from 29.0% to 32.8% on LLaMA-2-7B.

  7. Seeing Clearly by Layer Two: Enhancing Attention Heads to Alleviate Hallucination in LVLMs

    cs.CV 2024-11 conditional novelty 5.0 of 10

    A training-free intervention that broadcasts the densest vision-sink attention head's map to all heads in shallow layers lowers hallucination rates on CHAIR across several LVLMs.

  8. MCA-LLaVA: Manhattan Causal Attention for Reducing Hallucination in Large Vision-Language Models

    cs.CV 2025-07 conditional novelty 4.0 of 10

    MCA-LLaVA reindexes image tokens by sums of mirrored 2D coordinates so instruction tokens attend across the whole image, reducing hallucination on POPE, CHAIR, and MME.

  9. Probability Consistency in Large Language Models: Theoretical Foundations Meet Empirical Discrepancies

    cs.CL 2025-05 conditional novelty 4.0 of 10

    Perplexity is order-invariant by the chain rule for a true distribution, yet GPT-2 models trained on forward, backward, and permuted text deviate systematically, with permutation training producing the largest gaps.

Pith tools