TP-JSCC: Efficient Token Pruning and Joint Source-Channel Coding for LVLMs-based 3D-VQA in Cloud-Edge Collaboration

Zhengxin Li; Chongzhen Tian; Hui Yuan; Hao Liu; Sam Kwong

cs.CVOct 26, 2026

TP-JSCC: Efficient Token Pruning and Joint Source-Channel Coding for LVLMs-based 3D-VQA in Cloud-Edge Collaboration

Authors: Zhengxin Li, Chongzhen Tian, Hui Yuan, Hao Liu, Sam Kwong

Organizations: School of Control Science and Engineering, Shandong University, Jinan 250061, China · Key Laboratory of Machine Intelligence and System Control, Ministry of Education, Jinan 250061, China · Shandong Key Laboratory of Intelligent Communication and Sensing-Computing Integration, Jinan 250100, China · School of Computer and Control Engineering, Yantai University, Yantai, 264005, China ·

Abstract

Deploying large vision-language models (LVLMs) on resource-constrained embodied agents for 3D visual question answering necessitates cloud-edge collaboration. However, offloading massive visual data to the cloud over bandwidth-limited and noisy channels introduces severe latency and transmission distortion. To address this, we propose TP-JSCC, an efficient token pruning and joint source-channel coding method for LVLMs-based 3D-VQA in Cloud-Edge Collaboration. First, we develop a Local and Global Semantics-based Hybrid Token Pruning (LGS-HTP) module to significantly reduce the visual token count. By decoupling foreground-background tokens and jointly selecting local and global semantics, LGS-HTP effectively reduces redundancy while preserving reasoning-critical tokens. Second, we introduce a lightweight Token Joint Source-Channel Coding (T-JSCC) module to further compress the pruned tokens and ensure their robust transmission over error-prone channels. By employing lightweight cascaded linear layers and Signal-to-Noise Ratio (SNR)-adaptive mechanism, this module adapts to fluctuating channel conditions while maintaining minimal computational overhead. Extensive experiments on ScanQA and SQA3D datasets indicate the exceptional efficiency and robustness of our approach. Specifically, on ScanQA validation set, the LGS-HTP module could achieve 90% pruning ratio while incurring less than a 1% CIDEr loss. Building on this, the T-JSCC further compresses the tokens by 18 times, resulting in an overall 180 times data volume reduction for robust transmission, while saving 90% in computational complexity for LVLM deployment. Our code is available at https://github.com/yuanhui0325/TP-JSCC.

Figures & tables

Fig. 1: Framework of the proposed Token Pruning Joint Source-Channel Coding (TP-JSCC). The LGS-HTP module effectively selects global semantic tokens to reduce visual redundancy and achieve a scalable compression ratio. Simultaneously, the T-JSCC module encodes and transmits the selected sparse tokens to maintain transmission robust to complex wireless channel conditions.

Fig. 2: Illustration of the proposed LGS-HTP module for the $i$ -th view. The module comprises four steps: (1) inputs preprocessing for initial mask and token preparation, (2) enhancement for dynamic budget allocation, (3) clustering for semantic grouping, and (4) selection of the final retained tokens.

Fig. 3: Segmentation network for obtaining foreground-background priors.

Fig. 4: Pipeline of the hybrid token selection for the foreground tokens of the

i

-th view (the background tokens are selected following the same process).

Fig. 5: Overall architecture of the proposed T-JSCC network. A lightweight MLP network first reduces the token channel dimension to 64. The compressed tokens are then power-normalized and transmitted over error-prone channels. A symmetric MLP decoder is subsequently utilized for token reconstruction. Furthermore, the channel SNR

\gamma

is injected into both the encoding and decoding processes to achieve channel-adaptive transmission.

$\lambda_{\mathrm{bias}}$	Top-1 EM	CIDEr
1.0	28.5989	98.4296
1.1	28.5989	98.6438
1.2	28.4278	98.3734
1.3	28.5775	98.5151
1.4	28.5561	98.2152
1.5	28.6203	98.5228

TABLE I: Effect of the foreground-enhancement bias

\lambda_{\mathrm{bias}}

under 90% token pruning on the ScanQA validation set. The best, second-best, and third-best results are highlighted in bold , underline , and italic , respectively.

$K_{\mathrm{glo}}$	Top-1 EM	CIDEr
w/o Pruning	29.6898	99.4314
0	29.3262	98.1047
1	29.7540	99.5614
2	29.3048	99.5475
3	28.6845	97.4190
4	28.2353	96.9864

TABLE II: Effect of

K_{\mathrm{glo}}

in LGS-HTP under 80% token pruning on the ScanQA validation set. The best, second-best, and third-best results are highlighted in bold , underline , and italic , respectively.

$K_{\mathrm{glo}}$	Top-1 EM	CIDEr
w/o Pruning	29.6898	99.4314
0	28.2567	95.8569
1	28.6203	98.5228
2	27.8717	96.9990
3	28.0214	97.5871
4	28.0214	97.6235

TABLE III: Effect of

K_{\mathrm{glo}}

in LGS-HTP under 90% token pruning on the ScanQA validation set. The best, second-best, and third-best results are highlighted in bold , underline , and italic , respectively.

$D_{c}$	Top-1 EM	CIDEr
16	26.9733	92.6928
32	27.5722	94.5314
64	27.9144	96.3107
128	28.1711	96.5524
256	28.2995	97.5974

TABLE IV: Impact of the compressed channel dimension

D_{c}

in T-JSCC under 90% token pruning on the ScanQA validation set.

Fig. 6: Impact of the compressed channel dimension

D_{c}

on 3D-VQA performance under 90% pruning.

$\lambda_{\mathrm{N}}:\lambda_{\mathrm{P}}$	Top-1 EM	CIDEr
10:0	28.2567	95.0720
9:1	27.8289	95.2286
8:2	27.9358	95.4264
7:3	27.9144	95.2138
6:4	27.9572	96.0160
5:5	27.9144	96.3107

TABLE V: Effect of the T-JSCC loss-weight setting

\lambda_{\mathrm{N}}:\lambda_{\mathrm{P}}

on the ScanQA validation set. The best, second-best, and third-best results are highlighted in bold , underline , and italic , respectively.

SNR	Top-1 EM	CIDEr
w/o AWGN Channel	28.0428	97.3429
0	7.5508	25.3298
5	12.1497	38.1479
10	18.5882	65.4938
15	23.7647	87.1400
20	26.7380	94.0205

TABLE VI: 3D-VQA performance under different AWGN channel SNRs on the ScanQA validation set.

Pruning Method	Pruning Ratio	Top-1 EM	CIDEr
w/o Pruning	0%	29.6898	99.4314
FPS	80%	29.5187	94.0496
Random	80%	29.2406	99.3210
Uniform	80%	28.8770	101.1091
Proposed LGS-HTP	80%	29.7540	99.5614
FPS	90%	28.3850	89.2389

TABLE VII: Comparison of LGS-HTP with baseline token pruning methods at 80% and 90% pruning ratios on the ScanQA validation set. The best, second-best, and third-best results for each pruning ratio are highlighted in bold , underline , and italic , respectively.

Pruning Method	Pruning Ratio	ScanQA		SQA3D
Pruning Method	Pruning Ratio	Top-1 EM	CIDEr	Top-1 EM	CIDEr
w/o Pruning	0%	29.6898	99.4314	50.6394	132.6469
ToMe [39]	80%	27.8717	99.4191	48.8491	127.9719
GPrune [84]	80%	24.6203	77.9843	46.4052	120.9888
CDPrune [85]	80%	28.3209	91.8665	49.0480	128.19
DivPrune [86]	80%	29.1978	93.3695	50.0995	131.0206

TABLE VIII: Comparison with SOTA pruning methods on the ScanQA validation set and the SQA3D test set. The best, second-best, and third-best results for each pruning ratio are highlighted in bold , underline , and italic , respectively.

SNR	Top-1 EM	CIDEr
0	26.9733	94.0668
5	27.6578	96.0911
10	28.0428	96.4681
15	27.9786	97.1514
20	27.9572	97.0533
25	27.9572	96.5340

TABLE IX: End-to-end 3D-VQA performance of the proposed method under different SNRs with uniform SNR sampling (UFM) during training on the ScanQA validation set.

SNR	Top-1 EM	CIDEr
0	27.2941	94.8344
5	27.8503	96.6230
10	27.8717	96.2943
15	27.8717	96.5267
20	27.9786	96.8218
25	28.1070	96.9992

TABLE X: End-to-end 3D-VQA performance of the proposed method under different SNRs with progressive SNR range expansion (PRG) during training on the ScanQA validation set.

Token Decoupling	Selection Strategy	Top-1 EM	CIDEr
$\times$	FPS	28.5134	95.1825
$\times$	Random	28.0856	95.8969
$\times$	Uniform	28.1497	98.0843
✓	FPS	29.1551	95.0144
✓	Random	28.2995	96.0001
✓	Uniform	28.6203	98.5228

TABLE XI: Ablation study of LGS-HTP components: token decoupling and global semantic tokens selection strategies under 90% token pruning on the ScanQA validation set. The best, second-best, and third-best results are highlighted in bold , underline , and italic , respectively.

Guidance	Pruning Ratio	Top-1 EM	CIDEr
$\times$	80%	29.1551	98.4722
✓	80%	29.7540	99.5614
$\times$	90%	27.7219	97.3155
✓	90%	28.6203	98.5228

TABLE XII: Performance comparison between the unguided pruning variant and our proposed guidance strategy on the ScanQA validation set. The best results among the pruned variants for each pruning ratio are highlighted in bold .

TP-JSCC	Ratio	LVLM	LGS-HTP	T-JSCC	Total Saving
TP-JSCC	Compression	LVLM	LGS-HTP	T-JSCC	Total Saving
$\times$	$-$	93255.12	$-$	$-$	$-$
✓	$90\times$	19049.47	626.49	3.94	78.90%
✓	$180\times$	9659.90	626.49	1.92	88.97%

TABLE XIII: Computational complexity (in GFLOPs) of different modules and the LVLM under various compression ratios.

Visual Token Compression

Visual token compression reduces image and video representations to lower inference, memory, storage, or communication costs while retaining task-relevant information.

See all 270 papers

Seeing Once is Enough? Online Geometry-Aware Token Pruning for 3D Question Answering
Jul 5, 2026Ruei-Chi Lai, Bolivar Solarte, Chin-Hsuan Wu +2Efficient VLM Inference Multimodal QA
National Tsing Hua University · Industrial Technology Research Institute ITRI · University of Toronto +1
HiSC: Hierarchical Spatial Clustering Token Compression for Efficient 3D Scene Understanding
Aug 5, 2026Jiuhe Qu, Yingping Liang, Ying Fu3D Spatial Reasoning Visual Token Pruning
Beijing Institute of Technology Beijing, China
Don't Send What You Don't Need: Question-Guided Token Pruning as a Privacy Defense for Vision-Language Models
Sep 14, 2026Md Khalid Syfullah, Alvi Ataur KhalilVisual Question Answering Privacy Protection in VLMs
Transformative Innovation for Trustworthy AI and Network Security (TITANS) Lab, Computer Science, Southern Illinois University Carbondale, USA

Jul 5, 2026J/K move · Enter open · S save

Seeing Once is Enough? Online Geometry-Aware Token Pruning for 3D Question Answering

Ruei-Chi Lai, Bolivar Solarte, Chin-Hsuan Wu, Yi-Hsuan Tsai+1

National Tsing Hua University · Industrial Technology Research Institute ITRI · University of Toronto+1

Efficient VLM Inference Multimodal QA 3D VQA Visual Token Pruning 3D Scene Understanding

Open paper AskPDF arXiv

Recent Multi-modal Large Language Models (MLLMs) have demonstrated remarkable performance on 2D question answering tasks. However, extending these models to the 3D question answering remains challenging, as they typically require multiple views of the scene, which incurs substantial computational cost at inference. To mitigate this issue, existing solutions rely on strategic frame selection or token-merging algorithms that require preprocessing in advance all frames of the scene, i.e., an offline fashion. In contrast, we propose the first online token-pruning method that can be integrated seamlessly with current MLLM models for 3D question answering tasks, without additional training and with lower memory usage.Our key insight is to project each input frame into a shared voxel space using depth information and camera pose, identifying spatially-overlapped regions across frames and selectively pruning redundant image tokens before they enter the language model. Our method enables efficient online processing while reducing up to 50% of token usage. We apply this approach to Qwen2.5-VL-7B and Qwen3-VL-8B, demonstrating improved performance on the ScanQA, SQA3D, and OpenEQA-HM3D benchmarks.

Abstract

Visual Token Compression

Explore similar work

Seeing Once is Enough? Online Geometry-Aware Token Pruning for 3D Question Answering

HiSC: Hierarchical Spatial Clustering Token Compression for Efficient 3D Scene Understanding

Don't Send What You Don't Need: Question-Guided Token Pruning as a Privacy Defense for Vision-Language Models