VLN-Cache: Enabling Token Caching for VLN Models with Visual/Semantic Dynamics Awareness

Zihao Zheng; Zhihao Mao; Xingyue Zhou; Jiayu Chen; Maoliang Li; Xinhao Sun; Hailong Zou; Zhaobo Zhang

cs.ROMar 7, 2026

VLN-Cache: Enabling Token Caching for VLN Models with Visual/Semantic Dynamics Awareness

Authors: Zihao Zheng, Zhihao Mao, Xingyue Zhou, Jiayu Chen, Maoliang Li, Xinhao Sun, Hailong Zou, Zhaobo Zhang,

Organizations: School of Computer Science, Peking University, Beijing, China. · School of Computer Science, China University of Geosciences (Wuhan), Wuhan, China. · School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. · School of Electronics Engineering and Computer Science, Peking University, Beijing, China.

Abstract

Vision-and-Language Navigation (VLN) increasingly relies on large vision-language models, but their inference cost conflicts with real-time deployment. Token caching is a promising training-free strategy that avoids redundant computation by reusing stable visual tokens across frames. However, existing methods assume a static camera and fixed semantic focus, assumptions that VLN fundamentally violates. We identify two failure modes: (1) visual dynamics, where viewpoint shift displaces token positions across frames, causing position-wise matching to pair misaligned content; (2) semantic dynamics, where token relevance shifts across task stages as navigation progresses, making cached states stale. We propose VLN-Cache, a visual-dynamic-aware and semantic-dynamic-aware caching framework that introduces view-aligned remapping to recover geometric correspondences and a task-relevance saliency filter to veto reuse at semantic transitions. A layer-adaptive entropy policy further balances the per-layer reuse budget. Experiments on the R2R-CE simulation benchmark show up to 1.52x speedup while maintaining competitive navigation success rates.

Figures & tables

Fig. 1: Overview of the Proposed VLN-Cache Framework.

Fig. 2: Visual and semantic dynamics along a VLN trajectory.

Fig. 3: Position-wise token mismatch under viewpoint shift.

Fig. 4: Overview of the VLN-Cache framework.

Fig. 5: System implementation pipeline of VLN-Cache .

Fig. 6: Token reuse visualization on a staircase-to-bedroom navigation episode from R2R-CE.

Method	Backbone	NE $\downarrow$	OS $\uparrow$	SR $\uparrow$	SPL $\uparrow$
NaVid [ 27 ]	Vicuna-7B	5.47	49.1	37.4	35.9
MapNav [ 28 ]	Qwen2-7B	4.93	53.0	39.7	37.2
UniNaVid [ 26 ]	Vicuna-7B	5.58	53.3	47.0	42.7
NaVILA [ 8 ]	Llama-3-8B	5.22	62.5	54.0	49.0
StreamVLN [ 22 ]	Qwen2-7B	4.98	64.2	56.9	51.9
DualVLN [ 21 ]	Qwen2.5-VL-7B	4.05	70.7	64.3	58.5

TABLE I: Comparison with VLN methods on R2R-CE.

Fig. 7: Token reuse visualization on a bedroom-exit-to-rug navigation episode from R2R-CE.

Method	Latency $\downarrow$	Step	Episode	Token
	(ms/step)	Speedup	Speedup	Reuse
DualVLN [ 21 ]	637	1.00 $\times$	1.00 $\times$	0%
VLN-Cache (Ours)	419	1.52 $\times$	1.52 $\times$	31%

TABLE II: Inference efficiency comparison on R2R-CE.

Fig. 8: Sensitivity to $\tau_{v}$ and $\tau_{s}$ .

Fig. 9: Sensitivity to $k$ and $\rho_{\text{max}}$ .

Configuration	SR $\uparrow$	SPL $\uparrow$	Step Sp.	Token Reuse
w/o VLN-Cache (Baseline)	64.3	58.5	1.00 $\times$	0%
w/o View-Aligned Remap	62.4	56.8	1.41 $\times$	43%
w/o Semantic Gate	62.9	57.3	1.47 $\times$	30%
w/o Visual Gate	61.7	56.1	1.61 $\times$	49%
Full VLN-Cache (Ours)	63.1	57.6	1.52 $\times$	31%

TABLE III: Ablation study on R2R-CE.

Efficient VLM Inference

Efficient vision-language model inference reduces the compute, memory, latency, and energy needed to process images, video, and text.

See all 485 papers

FreqCache: Accelerating Embodied VLN Models with Adaptive Frequency-Guided Token Caching
Apr 27, 2026Zihao Zheng, Xingyue Zhou, Zhihao Mao +7Efficient VLM Inference Vision-Language Navigation
School of Computer Science, Peking University · School of Artificial Intelligence and Automation, Huazhong University of Science and Technology · School of Computer Science, China University of Geosciences (Wuhan) +3
DyGeoVLN: Infusing Dynamic Geometry Foundation Model into Vision-Language Navigation
Mar 22, 2026Xiangchen Liu, Hanghan Zheng, Jeil Jeong +5Vision-Language Navigation Efficient VLM Inference
KAIST · HKUST(GZ) · JD Explore Academy
Text-Vision Synergistic Token Caching: A Training-Free Framework for Efficient Vision-Language-Action Inference
Sep 28, 2026Qianer Li, Chengjie Zhang, Jingwen Chen +3Efficient VLA Model Inference Vision-Language-Action Models
College of Engineering, Southern University of Science and Technology (SUSTech), Shenzhen 518055, China · Master of Science in Data Science and Machine Learning, Department of Mathematics, Faculty of Science, National University of Singapore, Singapore

Apr 27, 2026J/K move · Enter open · S save

FreqCache: Accelerating Embodied VLN Models with Adaptive Frequency-Guided Token Caching

Zihao Zheng, Xingyue Zhou, Zhihao Mao, Songyu Sun+6

School of Computer Science, Peking University · School of Artificial Intelligence and Automation, Huazhong University of Science and Technology · School of Computer Science, China University of Geosciences (Wuhan)+3

Efficient VLM Inference Vision-Language Navigation

Open paper AskPDF arXiv

Vision-Language-Navigation (VLN) models exhibit excellent navigation accuracy but incur high computational overhead. Token caching has emerged as a promising training-free strategy to reduce this cost by reusing token computation results; however, existing token caching approaches rely on visual domain methods for cacheable token selection, leading to challenges when adapted to VLN models. 1) Visual domain methods become invalid when there is viewpoint migration. 2) Visual domain methods neglect critical edge information without the aid of additional algorithms. 3) Visual domain methods overlook the temporal variation of scenarios and lack adjustability in cache budgets. In this paper, we develop detailed analyses and find that the impacts of these challenges exhibit invariance and analyzability in the frequency domain. Based on these, we propose a frequency-guided token caching framework, called FreqCache. Utilizing the inherent properties of the frequency domain, FreqCache achieves optimal token cache establishment, refreshment, and adaptive adjustment. Experiments show that FreqCache achieves 1.59x speedup with ignorable overhead, showing the effect of integrating frequency domain methods in VLN token caching.

Abstract

Efficient VLM Inference

Explore similar work

FreqCache: Accelerating Embodied VLN Models with Adaptive Frequency-Guided Token Caching

DyGeoVLN: Infusing Dynamic Geometry Foundation Model into Vision-Language Navigation

Text-Vision Synergistic Token Caching: A Training-Free Framework for Efficient Vision-Language-Action Inference