See What You Are Told: Visual Attention Sink in Large Multimodal Models

• LMM decoder attention weights determine how much each text token focuses on corresponding visual information (Vaswani et al., 2017).Source: §1 · p. 1
• Irrelevant visual tokens receive high attention at fixed image locations across text tokens, arising from extreme activation in a few hidden dimensions and carrying limited image semantics.This resembles language-model attention sinks, which assign high weights to tokens with limited semantic meaning (Xiao et al., 2023; Sun et al., 2024a).Source: §1 · p. 2, Figure 1 · p. 2
• Removing visual sink tokens does not significantly affect response quality, despite their high attention weights.Source: §1 · p. 2
• VAR selects image-centric heads based on visual attention sinks, then redistributes the sinks’ excess attention budget to strengthen image focus in those heads.Source: §1 · p. 2
• VAR improves LMM performance across general vision-language, visual hallucination, and vision-centric tasks without additional training, models, or inference steps.Source: §1 · pp. 2–3

Visual attention in large multimodal models

• LMMs can overfocus on a few visual tokens, including tokens irrelevant to the text, while also failing to attend adequately to visual information overall.Visual contrastive decoding contrasts outputs with and without visual input, while other approaches increase attention weights assigned to images to encourage reliance on visual cues.Source: §2 · p. 3

Attention sink in language models

• Language-model attention sinks receive high weights despite limited semantic meaning and minimal predictive contribution, and are associated with massive activation of specific hidden-state dimensions.Similar behavior has been observed in vision transformers; prior work recalibrated sink-token attention in selected heads to improve language-model responses, and this paper extends the concept to LMMs as a visual attention sink.Source: §2 · p. 3

3 PRELIMINARIES

• An LMM combines a visual encoder and projector with a language model that receives system instructions, visual features, and text, then generates responses autoregressively.The paper calls the discrete inputs and their embeddings tokens.Source: §3 · p. 3
• Input tokens are partitioned into system, visual, and text indices and processed through transformer blocks containing multi-head attention and feed-forward networks.Source: §3 · p. 3
• The analysis focuses on attention weights from text tokens to visual tokens, representing their interactions as visual attention maps.Attention weights measure how much the model attends to input tokens; the figure illustrates that some irrelevant visual tokens have high attention and hidden-state activation.Source: §3 · p. 4, Figure 2 · p. 4

4 VISUAL ATTENTION SINK

• Irrelevant visual tokens receive high attention regardless of the text token and recur in fixed locations, leading the authors to identify their underlying behavior as a visual attention sink.The behavior stems from massive activation of specific hidden-state dimensions, analogous to attention sinks in language models.Source: §4 · p. 4

4.1 INVESTIGATING THE PROPERTY OF IRRELEVANT VISUAL TOKENS

• Irrelevant tokens show massive activation in specific dimensions also activated in BOS; among high-attention visual tokens, sink-dimension values separate into high and low groups.The authors define high-value tokens as visual sink tokens and set the analysis threshold to τ = 20; the sink dimensions are fixed by the base language model.Source: §4.1 · p. 5

4.2 ANALYSING THE CHARACTERISTICS OF VISUAL SINK TOKENS

• Masking visual sink tokens scarcely affects performance, unlike masking the same number of random visual tokens; sink tokens also contribute less to the text tokens’ residual stream.Visual sink tokens are mostly in less-informative backgrounds and share sink dimensions with text sink tokens, suggesting a common mechanism inherited from the base language model.Source: §4.2 · p. 6, Figure 3 · p. 6

4.3 SURPLUS ATTENTIONS IN VISUAL ATTENTION SINK: CAN WE RECYCLE THEM?

• Because sink attention carries little useful content while non-sink attention is closer to true image content, sink attention can be recycled as a budget to compensate for LMMs’ insufficient attention to images.The authors use the attention assigned to visual non-sink tokens to identify image-centric heads for subsequent attention redistribution.Source: §4.3 · pp. 6–7, Figure 4 · p. 7

5 VISUAL ATTENTION REDISTRIBUTION

• VAR selects image-centric heads using visual attention sinks, then reallocates attention from sink tokens to visual non-sink tokens only in those heads.The method has two steps: head selection and attention redistribution.Source: §5 · p. 7, Figure 5 · p. 8

5.1 SELECTING IMAGE-CENTRIC HEADS

• After excluding heads with less than 0.2 total visual attention, VAR selects heads whose visual non-sink ratio meets threshold ρ.The threshold controls how many heads are selected.Source: §5.1 · pp. 7–8, Figure 5 · p. 8
• Heads with high visual non-sink ratios concentrate on important visual tokens related to the text, while low-ratio heads show sparse, scattered attention.The paper reports this pattern from sorting and visualizing heads by their visual non-sink ratio.Source: §5.1 · pp. 7–8

5.2 REDISTRIBUTING ATTENTION WEIGHTS

• In selected heads, VAR transfers a portion p of sink-token attention to visual non-sink tokens according to their relative importance, preserving total attention at 1.The redistribution applies to all text tokens, including instructions and generated responses.Source: §5.2 · p. 8

6.1 EXPERIMENTAL SETTINGS

• Experiments apply VAR without additional training, models, or inference steps across six LMMs and general vision-language, visual hallucination, and vision-centric benchmark suites.The evaluated models include LLaVA-1.5, LLaVA-1.5-HD, VILA, Qwen2-VL, and InternVL2 variants.Source: §6.1 · pp. 8–9

6.2 MAIN RESULTS

• VAR reliably improves base-model performance across all three benchmark categories, and LLaVA-1.5-7B with VAR outperforms vanilla LLaVA-1.5-13B on GQA, VizWiz, MME, and POPE.The reported improvement holds across benchmarks without benchmark-specific hyperparameter tuning.Source: Table 1 · p. 9, Table 2 · p. 9, Table 3 · p. 9

6.3 ANALYSES AND DISCUSSIONS

• Attention-map analysis shows VAR redirects attention to key visual tokens, and the method can be incorporated into visual contrastive decoding.The authors conclude that steering internal attention can improve multimodal capability.Source: Figure 6 · p. 10, §6.3 · p. 10
• Redistributing attention across all heads can yield 0.0 scores, while redistributing only to visual tokens is most effective; text-only redistribution yields little improvement or can reduce performance.Redistribution to both visual and text tokens slightly improves performance, but selecting image-centric heads is crucial.Source: §6.3 · p. 10
• VAR improves performance across reasonable hyperparameter ranges; τ and p have consistent optimal values across tasks, while optimal ρ is consistent across benchmarks within each task type.The authors use one τ and p across tasks and one ρ per task type.Source: §6.3 · p. 10

7 CONCLUSION

• VAR recycles surplus attention from irrelevant visual sink tokens to emphasize image information relevant to text, improving LMM visual processing through attention-map edits alone.The authors present this as a way to improve multimodal capabilities without additional training.Source: §7 · p. 10
• LMMs consistently attend to irrelevant parts of images, which the paper identifies as visual attention sinks.The paper investigates the properties and characteristics of these tokens.Source: §7 · p. 10

A.1 MASSIVE ACTIVATION OF SINK DIMENSIONS

• Visual sink tokens show massive activation in the same hidden-state dimensions as BOS tokens across LMMs with varied architectures, training schemes, and scales.The authors interpret this shared pattern as an innate LMM property resembling language-model sink tokens.Source: Figure 7 · p. 16, §A.1 · p. 16
• Sink dimensions are consistent across layers and heads, agnostic to input tokens, and unchanged by fine-tuning, consistent with an innate property arising during pre-training.The passage grounds these background properties in Sun et al. (2024a) and attributes the pre-training account to Darcet et al. (2023) and Gu et al. (2024).Source: §A.1 · p. 16
• LLaVA-1.5-7B retains LLaMA2-7B’s sink dimensions D_sink={1415, 2533} after multimodal fine-tuning.The analysis sets sink dimensions from each LMM’s pretrained language model: {2100, 4743} for LLaVA-1.5-13B, LLaVA-1.5-HD-13B, and VILA-13B; {458, 2570} for Qwen2-VL-7B; and {3584} for InternVL2-8B.Source: §A.1 · p. 16
• The visual sink tokens share characteristics with language-model sink tokens, including massive activation in matching dimensions across the LMMs examined.The paper reports this pattern across models with different architectures, training schemes, and scales.Source: §A.1 · p. 16

Location of visual sink tokens

• Visual sink tokens are more likely than other visual tokens to fall in background regions, supporting their disconnection from the image’s main subject.The analysis treats regions outside the main object as background; the finding resembles sink-token patterns in ViT (Darcet et al., 2023) and language models (Ferrando & Voita, 2024; Xiao et al., 2023; Yu et al., 2024).Source: §A.2 · p. 17, Table 6 · p. 17

Layer-wise analysis

• Massive hidden-state activation of visual sink tokens emerges in early layers, and the tokens persist through the last layer.On LLaVA-1.5-7B, activation appears by layer 2; early FFNs may write the activation, which residual connections preserve in subsequent layers, paralleling language-model sink behavior (Sun et al., 2024a; Cancedda, 2024).Source: §A.2 · p. 17, Figure 8 · p. 17

Further discussions and future works

• How LMMs select visual sink tokens and how the sinks emerge during multimodal training remain open questions, although they may be inherited from base language models.Sink dimensions in multimodal models are identical to those in their base language models; the tokens are also described as semantically less meaningful and background-located. Similar sinks emerge in vision and language models after sufficient training on massive datasets (Darcet et al., 2023; Gu et al., 2024).Source: §A.2 · pp. 17–18

A.3 ANALYSIS ON IMAGE-CENTRIC HEADS

• Ablating image-centric heads causes larger performance drops than ablating random heads across ρ settings, indicating their importance for visual processing.The study masks attention from visual tokens to text tokens and varies ρ, which controls the number of selected heads.Source: §A.3 · p. 18
• Ablating complementary heads leaves performance relatively stable across ρ, suggesting those heads are not responsible for visual information processing.Source: §A.3 · p. 18
• Text tokens that require visual information have a larger portion of image-centric heads, and VAR adjusts emphasis according to token context.Source: §A.3 · p. 18, Figure 10 · p. 19
• For example, the visually grounded token “Sub” has more image-centric heads than “in,” which can be inferred from the question alone.Source: §A.3 · p. 18

B MORE EXPERIMENTS ON VISUAL ATTENTION REDISTRIBUTION

• VCD uses contrastive decoding between original and distorted visual inputs but requires two models and extra inference steps; VAR can be incorporated into VCD without those extra steps.The combined method contrasts outputs from a VAR model and a model given distorted visual input.Source: §B.1 · p. 19

B.1 VISUAL CONTRASTIVE DECODING: A CASE STUDY ON INCORPORATING VISUAL ATTENTION REDISTRIBUTION INTO OTHER METHODS

• VCD with VAR improves VCD performance on MME, POPE, and GQA.The evaluation used LLaVA-1.5-7B; the supplied passages do not provide the benchmark scores.Source: §B.1 · p. 19, Table 7 · p. 19

Determination of hyperparameters

• Optimal hyperparameters transfer across benchmarks within each task type; τ and p are consistent across tasks, while ρ varies by task.Tuning uses a random 10% pseudo-validation sample, and the experiments set τ = 20 and p = 0.6; ρ is 0.8 for general vision-language, 0.5 for visual hallucination, and 0.9 for vision-centric tasks.Source: §B.2 · pp. 19–20, Figure 11 · p. 20

Impact of hyperparameters

• VAR consistently improves the baseline across tasks within a reasonable range of hyperparameter values, showing robustness to hyperparameter choice.The impact is evaluated on representative general vision-language, visual hallucination, and vision-centric benchmarks.Source: §B.2 · pp. 20–21, Figure 12 · p. 21
• τ = 20 gives the best performance and separates irrelevant from relevant visual tokens clearly.τ is the threshold for selecting sink tokens.Source: §B.2 · p. 21
• The suitable ρ varies by task, but VAR robustly improves performance for ρ ≥ 0.5.A lower ρ selects more image-centric heads for hallucination reduction, while general vision-language tasks use a higher ρ to select fewer such heads.Source: §B.2 · p. 21
• p = 0.6 is best across benchmarks, although other values can also improve performance.p controls the portion of sink-token attention weights returned to the attention budget.Source: §B.2 · p. 21

C LIMITATIONS AND FUTURE WORKS

• Raw visual attention maps may mislocalize objects when the model or encoder misinterprets images, motivating localization refinement and evaluation on different, larger LMMs.The authors also propose studying how visual sink tokens are learned during multimodal training.Source: C LIMITATIONS AND FUTURE WORKS · p. 22

D.1 DETAILS OF BENCHMARKS

• General vision-language evaluation covered VQAv2, GQA, VizWiz, ScienceQA, and TextVQA, with GQA test-dev-balanced and VizWiz and TextVQA validation splits.ScienceQA comprises 21,208 multimodal multiple-choice science questions; the passage describes its image, text, and combined contexts and curriculum-derived subject coverage.Source: §D.1 · p. 22
• Other general-task evaluations used MME perception tasks, English-only MMBench with circular evaluation, and the image-based SEED-Bench subset.LLaVA-Bench (In-the-Wild) and MM-Vet were also included, with LLaVA-Bench used for qualitative case studies.Source: §D.1 · pp. 22–23
• CHAIR evaluated captioning object hallucinations using CHAIRI, the share of mentioned objects that are hallucinated, and CHAIRS, the share of captions containing hallucinations.The evaluation also included VQA hallucination assessment.Source: §D.1 · p. 23
• POPE assessed object hallucination with balanced present and absent-object queries and Accuracy, Precision, Recall, and F1; MMHal-Bench used GPT-4 ratings on 96 OpenImages questions.POPE draws 27,000 query-answer pairs from 500 images per dataset and tests random, popular, and adversarial negative-sample strategies.Source: §D.1 · p. 23
• Vision-centric evaluation used MMVP’s 300 image-question examples and CV-Bench’s 2,638 examples covering 2D and 3D understanding in multiple-choice format.MMVP contains CLIP-blind image pairs; CV-Bench repurposes datasets including ADE20k, COCO, and OMNI3D.Source: §D.1 · p. 23

D.2 DETAILS OF EVALUATION SETTING

• Experiments used inference only, without training, on a single NVIDIA GeForce RTX A6000 48GB GPU.Source: §D.2 · p. 23
• GPT-4-based evaluations for LLaVA-Bench, MM-Vet, and MMHal-Bench used gpt-4-0613 with a maximum output length of 1024.Source: §D.2 · p. 24

Investigation of Sink Dimension Values

• A scatter plot relates sink-dimension values to visual-token attention weights across 100 MME visual question-answering samples, with 10 samples selected from each category.The plot is shown in Fig. 3(a).Source: §D.3 · p. 24

Token Masking Experiment

• Sink-token attention knockout masks attention from tokens with sink-dimension values at least 20 to text tokens; random knockout masks the same number of tokens for comparison.The LLaVA-1.5-7B model is evaluated on POPE, with results shown in Fig. 3(b); random tokens are selected independently at each layer.Source: §D.3 · p. 24
• Attention knockout may distort model behavior by forcing masked attention weights to zero, while independently selected random tokens can be inconsistent across layers.The authors nonetheless regard the masking experiments as informative about visual sink tokens’ influence on performance.Source: §D.3 · p. 24

Contribution Analysis

• Visual sink-token attribution contributions are measured against the average for all visual tokens using LLaVA-1.5-7B on POPE, with results shown in Fig. 3(c).The analysis is described as measuring the average value of the visual sink tokens’ attribution contributions.Source: §D.3 · p. 24

E ADDITIONAL QUALITATIVE RESULTS

• Visual sink tokens consistently receive high attention across text tokens, regardless of their content.The visualized tokens include both instructions and generated responses.Source: E ADDITIONAL QUALITATIVE RESULTS · p. 25, Figure 13 · p. 26
• With VAR, LLaVA-1.5-7B generates more accurate, image-relevant responses than without VAR.The examples contrast incorrect descriptions from vanilla LLaVA-1.5-7B with responses that incorporate visual information more effectively after VAR.Source: E ADDITIONAL QUALITATIVE RESULTS · p. 25, Figure 14 · p. 27
• Figure 15 illustrates a qualitative GPT-4 evaluation example for open-ended generation tasks, with Assistant 1 and Assistant 2 referring respectively to GPT-4 and LLaVA-1.5-7B + VAR.Source: Figure 15 · p. 28