Representation Dynamics Reveal Semantic Saliency and Similarity for Visual Token Pruning in MLLMs

Weixuan Li; Zikun Zhou; Xinyi Zhuang; Xinyan Guo; Rui Tian; Chuyao Zhang; Lin Gao

cs.CVSep 29, 2026

Representation Dynamics Reveal Semantic Saliency and Similarity for Visual Token Pruning in MLLMs

Authors: Weixuan Li, Zikun Zhou, Xinyi Zhuang, Xinyan Guo, Rui Tian, Chuyao Zhang, Lin Gao

Organizations: Harbin Institute of Technology, Shenzhen · Shenzhen Loop Area Institute

Abstract

Multimodal large language models (MLLMs) incur high inference latency from long visual token sequences. Existing pruning methods commonly use attention maps or output features to estimate token importance or redundancy. Several recent approaches also exploit representation changes, but when and how these changes reflect foreground saliency and semantic consistency remain insufficiently understood. We analyze visual token representation dynamics across encoder depth and uncover two findings. First, the relationship between token update magnitudes and foreground saliency is layer-dependent: large token updates concentrate on foreground regions in two depth intervals, separated by several sink-dominated layers at intermediate depths. Second, similarities between token update directions better distinguish same-class from different-class tokens than those between encoder output features. Building on these findings, we propose MSDG-Prune, a training-free method that uses update magnitudes and directions to preserve salient and diverse visual information. Specifically, we group tokens by update-direction similarity and use query-weighted saliency derived from update magnitudes across a chosen depth window for group-wise token pruning. Extensive experiments across four MLLMs demonstrate the effectiveness and generalizability of MSDG-Prune. On LLaVA-NeXT, it retains 91.9% of uncompressed performance on average with only 5.6% of visual tokens, while achieving a 7.8x prefilling speedup. Code is available at https://github.com/liweixuan-hitsz/MSDG-Prune.

Figures & tables

Figure 1: Output features (OF) versus update directions (UD) for semantic similarity in LLaVA-1.5. Panels (a) and (b) visualize the top 40 tokens ranked by cosine similarity to the airplane anchor token ( red ), using OF and UD, respectively. Panels (c) and (d) show cosine similarity distributions on COCO 2017 computed using OF and UD, respectively. For each anchor token, we compute cosine similarities to other tokens and group the resulting pairs by whether they share the same semantic class as the anchor token. Blue and gray curves represent same-class and different-class pairs, respectively. The horizontal axis indicates cosine similarity, and the vertical axis shows the percentage of pairs within each similarity bin. Compared with OF, UD shows less overlap between the two groups, indicating better semantic discrimination.

Figure 2: Layer-wise token update magnitudes in the vision encoder of LLaVA-1.5. Each map overlays patch-token update magnitudes across a Transformer block on the input image. L00-01 denotes the change from token states entering Layer 0 to those entering Layer 1. Because LLaVA-1.5 uses the penultimate-layer output of its CLIP ViT-L/14@336 ( Radford et al., 2021 ; Liu et al., 2024a ) vision encoder as input to the visual projector, the visualization ends at L22-23 and omits L23-24. The maps illustrate depth-dependent spatial patterns in token updates, including foreground enhancement before and after the sink-dominated stage.

Figure 3: Layer-wise analysis of token update magnitudes in LLaVA-1.5 on COCO 2017 validation images. Top: foreground recall of the top

30\%

of tokens ranked by update magnitude, with random selection as the dashed baseline. Bottom: the max-to-median update magnitude ratio,

\max_{i}v_{i}^{\ell}/\operatorname{median}_{i}v_{i}^{\ell}

, with a dashed diagnostic threshold of

10

. Both metrics are computed per image and then averaged across images. The highlighted Layers 11–12 exhibit sharp ratio peaks, marking the sink-dominated stage in the corresponding vision encoder.

Figure 4: Overview of MSDG-Prune. After sink filtering, update magnitudes over a selected foreground-enhanced window provide visual saliency scores, while update-direction similarity organizes tokens into semantic groups. Visual saliency and query relevance jointly guide budget allocation across groups and token selection within each group.

Method	GQA	MMB	MME	POPE	SQA	VQA ${}^{\text{v2}}$	VQA ${}^{\text{Text}}$	SEED ${}^{\text{I}}$	VizWiz	RelAcc.
Uncompressed baseline (100%)
Vanilla CVPR 2024	61.9	64.7	1862	85.9	69.5	78.5	58.2	60.5	54.3	100%
Token retention: 33.3%
FastV ECCV 2024	52.7	61.2	1612	64.8	67.3	67.1	52.5	57.1	50.8	89.1%
SparseVLM ICML 2025	57.6	62.5	1721	83.6	69.1	75.6	56.1	55.8	50.5	95.2%
DART EMNLP 2025	60.0	63.6	1856	82.8	69.8	76.7	57.4	51.5	54.9	97.1%

Table 1: Performance comparison on LLaVA-1.5-7B. Vanilla refers to the uncompressed baseline. The subscript after each method gives the publication venue.

Table 6

Method	Tokens	Inference Time $\downarrow$	Prefilling Time $\downarrow$	POPE (F1) $\uparrow$
LLaVA-NeXT	2880	254 ms	218 ms	86.4
SparseVLM	160	199 ms	119 ms	69.3
VisionZip	160	84 ms	27.8 ms	74.8
PruneSID	160	89 ms	27.8 ms	76.9
Ours	160	96 ms	27.8 ms	87.1

Table 5: Inference latency and POPE F1 score on LLaVA-NeXT-7B. Latencies are averaged per sample using a single NVIDIA A800-80 GB GPU.

Method	Retain 33.3%					Retain 22.2%					Retain 11.1%
Method	GQA	MME	POPE	SQA	RelAcc.	GQA	MME	POPE	SQA	RelAcc.	GQA	MME	POPE	SQA	RelAcc.
Ours	58.5	2332	84.7	87.8	98.5%	57.7	2297	84.0	87.3	97.4%	55.4	2178	81.3	86.3	94.1%
Output features	58.1	2286	84.0	87.5	97.5%	56.9	2254	82.8	87.1	96.2%	54.8	2178	80.1	86.0	93.5%
Attention	57.9	2319	86.1	87.9	98.5%	56.9	2268	84.4	87.6	97.0%	53.8	2176	81.3	86.5	93.5%
Accumulated updates	58.5	2317	84.4	87.9	98.3%	56.5	2272	83.2	87.3	96.4%	54.3	2129	80.1	85.9	92.7%
Keep sink tokens	58.4	2330	84.7	87.7	98.4%	57.1	2264	83.1	87.2	96.5%	54.8	2143	81.2	86.3	93.5%

Table 6: Component ablations and staged token selection on Qwen2.5-VL-7B. Ours is the full method; Output features uses encoder-output features for grouping. Query (global) ranks by query relevance without grouping; Query

\times

Visual (global) ranks by the product of query relevance and visual saliency, still without grouping.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Algorithm 1 MSDG-Prune Visual Token Selection
Require: encoder states $\{h_{i}^{\ell}:1\leq i\leq N,\ 0\leq\ell\leq L\}$ , projected visual tokens $\{\bm{x}_{i}\}_{i=1}^{N}$ , where $N$ is the number of visual tokens before pruning and $L$ is the number of encoder layers, query embeddings $\{\bm{t}_{j}\}_{j\in\mathcal{Q}}$ , where $\mathcal{Q}$ indexes the query tokens, budget $B$ , group count $K$ , $\ell_{0}$ and $\ell_{d}$ (start and end layers of the update direction $d_{i}$ ), $\ell_{\star}$ (peak concentration layer; Appendix C.1 ), $\ell_{s}$ and $\ell_{e}$ (start and end of the saliency window for the visual saliency score), diagnostic coordinate $d_{s}$ , threshold $\tau$ , and integer initialization seed $s$ ; $\operatorname{unit}(z)=z/\max\{\lVert z\rVert,10^{-12}\}$ . $\operatorname{TopIdx}_{b}$ returns the indices of the $b$ largest scores, breaking ties by original token order.
1:	$\mathcal{V}\leftarrow\{i:\|(h_{i}^{\ell_{\star}+1})_{d_{s}}\|\leq\tau\}$ ; require $B\leq\|\mathcal{V}\|$ , $1\leq K\leq\|\mathcal{V}\|$ , and $\mathcal{Q}\neq\varnothing$
2:	for each token $i\in\mathcal{V}$ do
3:	$\Delta h_{i}\leftarrow\operatorname{unit}(h_{i}^{\ell_{d}})-\operatorname{unit}(h_{i}^{\ell_{0}})$ ; $d_{i}\leftarrow\operatorname{unit}(\Delta h_{i})$
4:	$u_{i}\leftarrow\lVert h_{i}^{\ell_{e}}-h_{i}^{\ell_{s}}\rVert$ ; $\alpha_{i}\leftarrow\max_{j\in\mathcal{Q}}\operatorname{unit}(\bm{x}_{i})^{\top}\operatorname{unit}(\bm{t}_{j})$
5:	end for

Appendix

Table 9

Panel of Figure 6 — Figure 6: Cross-image sink signature on COCO 2017 validation images on LLaVA-NeXT and Qwen2.5-VL. Top: image-mean layer-wise foreground recall when selecting the top $30\%$ of tokens by $v_{i}^{\ell}$ . Bottom: mean concentration ratio $\max_{i}v_{i}^{\ell}/\operatorname{median}_{i}v_{i}^{\ell}$ ; the dashed line marks the threshold ( $5$ on LLaVA-NeXT and $10$ on Qwen2.5-VL). A ratio spike at a fixed intermediate depth indicates isolated tokens with large update magnitudes, consistent with the sink-dominated stage.

Figure 7: Image-mean Foreground Recall@30% for windows of width

w=5

on COCO 2017 validation images. At each input-state start

\ell_{s}

, patches are ranked by

\lVert h_{i}^{\ell_{s}+5}-h_{i}^{\ell_{s}}\rVert

. Horizontal dashed lines show the random selection expectation (approximately

30\%

); vertical dotted lines mark the default starts used in the main experiments.

Figure 8: Spatial maps of update magnitudes $v_{i}^{\ell}$ across LLaVA-1.5 and Mini-Gemini vision encoder layers for the tennis scene.

Figure 9: Spatial maps of update magnitudes $v_{i}^{\ell}$ across LLaVA-NeXT vision encoder layers for the tennis scene.

Panel of Figure 10 — Figure 10: Spatial maps of update magnitudes $v_{i}^{\ell}$ across LLaVA-NeXT vision encoder layers for the kite scene.

Figure 12: Spatial maps of update magnitudes $v_{i}^{\ell}$ across Qwen2.5-VL vision encoder layers for the tennis scene.

Method	GQA	MMB	MME	POPE	SQA	VQA ${}^{\text{v2}}$	VQA ${}^{\text{Text}}$	SEED ${}^{\text{I}}$	VizWiz	RelAcc.
Uncompressed baseline (100%)
Vanilla	63.2	67.7	1818	85.9	72.8	80.0	61.3	66.9	56.6	100%
Token retention: 33.3%
VisionZip	59.6	65.9	1770	86.4	72.8	78.0	58.6	65.2	54.9	97.5%
PruneSID	59.6	65.9	1770	86.4	72.8	78.0	58.6	65.2	56.0	97.7%
Ours	60.1	67.4	1801	86.1	73.7	76.8	59.4	65.3	56.0	98.3%

Appendix

Table 9: Performance comparison on LLaVA-1.5-13B. RelAcc. averages the relative scores across the nine listed benchmarks.

Method	GQA	MMB	MME	POPE	SQA	VQA ${}^{\text{v2}}$	VQA ${}^{\text{Text}}$	SEED ${}^{\text{I}}$	VizWiz	RelAcc.
Uncompressed baseline (100%)
Vanilla	65.4	70.0	1858	86.2	73.5	81.8	64.3	71.9	64.0	100%
Token retention: 22.2%
PruneSID	62.4	67.0	1817	85.6	70.1	79.1	60.2	68.5	60.2	95.9%
Ours	63.6	66.6	1820	87.0	72.8	78.6	62.0	68.5	60.1	96.9%
Token retention: 11.1%

Appendix

Table 10: Performance comparison on LLaVA-NeXT-13B.

Method	GQA	MMB	MME	POPE	SQA	MMMU	VQA ${}^{\text{Text}}$	RelAcc.
Uncompressed baseline (100%)
Vanilla	59.3	86.3	2428	84.3	91.5	61.1	76.5	100%
Token retention: 33.3%
VisionZip	57.3	85.1	2307	82.1	88.9	57.7	72.1	96.2%
PruneSID	57.0	81.6	2209	82.7	85.4	57.6	70.6	94.2%
Ours	56.4	84.4	2396	81.6	89.1	58.3	72.9	96.6%

Appendix

Table 11: Performance comparison on Qwen2.5-VL-32B.

Method	Retain 33.3%					Retain 22.2%					Retain 11.1%
Method	GQA	MME	POPE	SQA	RelAcc.	GQA	MME	POPE	SQA	RelAcc.	GQA	MME	POPE	SQA	RelAcc.
Ours	60.1	1806	86.9	68.7	98.5%	58.9	1778	86.7	68.7	97.6%	57.7	1708	85.6	68.4	95.8%
Grouping
Output features	60.0	1771	86.7	68.4	97.8%	58.9	1712	86.6	68.7	96.7%	57.8	1680	84.4	68.4	95.1%
Saliency scoring
Attention	58.7	1795	84.3	68.7	97.1%	57.9	1775	83.3	68.4	96.1%	56.4	1675	81.2	68.7	93.6%

Appendix

Table 12: Grouping, saliency scoring, and update aggregation on LLaVA-1.5-7B. Ours uses unit update directions and endpoint displacement. Output features replaces the grouping representation with

h_{i}^{o}

. Attention replaces the visual saliency score

u_{i}

with [CLS] attention from the output layer. Accumulated replaces endpoint displacement with accumulated update magnitude.

$\ell_{0}$	GQA	MMB	MME	POPE	SQA	RelAcc.	GQA	MMB	MME	POPE	SQA	RelAcc.	GQA	MMB	MME	POPE	SQA	RelAcc.	Avg.
	Retain 33.3%						Retain 22.2%						Retain 11.1%
$0$	58.1	81.7	2323	84.9	88.0	98.2%	57.0	80.2	2300	83.7	87.1	96.8%	54.5	77.6	2165	81.3	86.4	93.4%	96.1%
$1$	58.2	82.0	2318	84.7	87.7	98.1%	57.3	80.8	2290	84.1	87.3	97.1%	54.8	77.7	2212	81.4	86.3	94.0%	96.4%
$2$	58.5	82.6	2332	84.7	87.8	98.5%	57.7	81.2	2297	84.0	87.3	97.3%	55.4	78.4	2178	81.3	86.3	94.0%	96.6%
$3$	58.2	82.0	2329	85.0	87.8	98.3%	57.5	80.7	2294	84.6	87.0	97.1%	55.4	77.1	2161	81.3	86.4	93.6%	96.3%

Appendix

Table 13: Ablation of the early endpoint

\ell_{0}

on Qwen2.5-VL-7B. Avg. is the mean RelAcc. over the three retention ratios.

$K$	GQA	MMB	MME	POPE	SQA	RelAcc.	GQA	MMB	MME	POPE	SQA	RelAcc.	GQA	MMB	MME	POPE	SQA	RelAcc.	Avg.
	Retain 33.3%						Retain 22.2%						Retain 11.1%
$8$	57.6	81.2	2276	83.3	87.6	97.0%	56.2	79.8	2263	81.8	86.7	95.5%	53.7	76.7	2105	78.8	85.3	91.6%	94.7%
$12$	57.8	80.8	2286	83.6	87.7	97.1%	56.2	79.7	2249	82.7	86.8	95.6%	54.4	75.8	2131	80.0	85.6	92.2%	95.0%
$16$	58.0	81.3	2302	84.4	87.8	97.7%	56.8	79.7	2230	82.5	87.2	95.7%	54.6	76.6	2130	80.8	85.7	92.6%	95.3%
$20$	58.5	82.6	2332	84.7	87.8	98.5%	57.7	81.2	2297	84.0	87.3	97.3%	55.4	78.4	2178	81.3	86.3	94.0%	96.6%
$24$	58.5	81.9	2361	85.3	87.8	98.7%	57.5	81.1	2311	83.8	87.3	97.3%	55.3	78.0	2170	81.4	86.3	93.8%	96.6%

Appendix

Table 15: Ablation of the group count

K

on Qwen2.5-VL.

Table 23

Uncompressed baseline (100%)
Window	GQA	MME	POPE	SQA	RelAcc.
Vanilla	60.9	2310	86.3	88.9	100%
Token retention: 22.2%
$19{\to}22$ ( $w{=}3$ )	57.8	2252	84.2	87.1	97.0%
$19{\to}24$ ( $w{=}5$ )	57.7	2297	84.0	87.3	97.4%
$19{\to}26$ ( $w{=}7$ )	57.6	2273	84.0	87.4	97.2%

Appendix

Table 18: Saliency window width ablation on Qwen2.5-VL-7B with the start fixed at Layer

19

and all other components unchanged. An interval

a{\to}b

covers updates from Layer

a

through Layer

b-1

, so varying

b

changes the width while preserving the starting layer.

Token retention: 22.2%
Method	GQA	MME	POPE	SQA	RelAcc.
$\alpha$ Alloc., $u$ Rank	57.4	2307	84.2	87.1	97.4%
$u$ Alloc., $\alpha$ Rank	55.0	2264	81.1	86.6	94.9%
Ours	57.7	2297	84.0	87.3	97.4%
Token retention: 11.1%
$\alpha$ Alloc., $u$ Rank	55.0	2166	81.5	86.5	94.0%

Appendix

Table 19: Ablation of signal assignment on Qwen2.5-VL-7B. Allocation and ranking refer to budget allocation across groups and token selection within each group, respectively.

Figure 16: Failure cases of MSDG-Prune on Qwen2.5-VL-7B at 11.1% target token retention. Each case pairs the original image (left) with the retained-token visualization (right). GT denotes the ground-truth answer and Pred denotes the answer after pruning; the uncompressed model answers both questions correctly. Discarded patches are covered by a translucent gray mask.

Figure 17: Qualitative comparison at a retention ratio of 11.1%. Columns show the original image and the tokens retained by VisionZip, PruneSID, and MSDG-Prune, with the corresponding query below each example. Discarded patches are covered by a translucent gray mask.

Explore similar work

S $^2$ Prune: Spatially Structured Visual Token Pruning for Multimodal Large Language Models
Sep 1, 2026Yuanyuan Jia, Shunpu Tang, Qianqian YangVisual Token Pruning Multimodal Large Language Models
College of Information Science and Electronic Engineering, Zhejiang University Hangzhou, China
TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient MLLM Inference
Jun 25, 2026Tinghao Wang, Yichen Guo, Rui Huang +11Visual Token Pruning Multimodal Large Language Models
State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University · Nanyang Technological University · University of Electronic Science and Technology of China +1
SinkPruner: Sink-Free Visual Token Pruning for Multimodal Large Language Models
Sep 1, 2026Shiyu Li, Zi-Yuan Hu, Shijia Huang +3Visual Token Pruning Long Visual-Token Sequences
1The Chinese University of Hong Kong · 2Weitu AI · 3Peking University

Sep 1, 2026, cs.CVJ/K move · Enter open · S save

S $^2$ Prune: Spatially Structured Visual Token Pruning for Multimodal Large Language Models

Yuanyuan Jia, Shunpu Tang, Qianqian Yang

College of Information Science and Electronic Engineering, Zhejiang University Hangzhou, China

Visual Token Pruning Multimodal Large Language Models Geometric Bias

Open paper AskPDF arXiv

Visual token pruning reduces the inference overhead of multimodal large language models (MLLMs) by retaining only a subset of visual tokens. Existing methods usually select tokens based on importance or redundancy. However, we observe that these criteria produce stable spatial biases across inputs and do not always outperform simple Uniform Grid sampling, highlighting the value of broad spatial coverage. Motivated by this, we propose S

^2

Prune, a training-free pruning method that preserves spatial coverage while adapting token density to local image structure. We first divide the image into regions and assign at least one token to each region to preserve coverage. The remaining token budget is then distributed according to Laplacian variation, giving more tokens to regions with richer structure. We then use Early Representation Change (ERC), computed from the first decoder block, to select representative tokens within each region. We evaluate S

^2

Prune across diverse settings and two MLLM architectures. On Qwen2.5-VL-7B-Instruct, it achieves the highest average accuracy among the evaluated training-free pruning methods. With only 32 of the original 576 visual tokens, it still retains 79.3% of the full-model performance. Code is available at https://github.com/yuanyuanjia71-spec/S2Prune.

Abstract

Explore similar work

S2^22Prune: Spatially Structured Visual Token Pruning for Multimodal Large Language Models

TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient MLLM Inference

SinkPruner: Sink-Free Visual Token Pruning for Multimodal Large Language Models

S $^2$ Prune: Spatially Structured Visual Token Pruning for Multimodal Large Language Models