Foundation models like V-JEPA are increasingly deployed across a heterogeneous edge-cloud continuum spanning smart glasses, phones, robots, and servers. Yet, they are typically trained as rigid, fixed-capacity networks bound to one device, supporting neither reconfigurable compute budgets nor collaborative execution. Existing approaches address only one axis at a time: slimmable and NAS-based methods provide width elasticity but assume single-device inference, whereas split-computing methods enable cross-device collaboration only for fixed partitions and static capacities. We introduce Elastic V-JEPA (EV-JEPA), a single model trained once and deployable everywhere along the continuum. Its width is reconfigurable, and it is natively partitionable across heterogeneous devices. At its core is Transition-Aware Triplet Training (T3). At each step we pick a narrow and a wide width, and train three paths: each width alone, plus a hybrid path that runs the early layers narrow and the rest wide. The hybrid switch point is where one device hands off to another, so the same weights learn both single-device and cross-device execution. Across various benchmarks, EV-JEPA falls within 1 accuracy point of independently trained baselines while improving over previous elastic training approaches by up to 11 accuracy points. Narrow-to-wide cross-device hybrid execution recovers to within 1.1 points of the highest-capacity standalone model and supports many-to-one, any-to-any, and variable-split configurations. Finally, we show that the elastic principle extends beyond the backbone to an elastic token compression mechanism that exposes controllable trade-offs among on-device computation, latency, and accuracy.
Figures & tables
Figure 1: One elastic model for the device continuum. We propose a single elastic network that is trained once and deployed across the entire continuum: (1) its width is reconfigurable at inference time, allowing the same parameters to serve any compute budget (Standalone Inference) ; and (2) its computation can be split between devices, with early layers executed on-device and remaining layers on a stronger device, communicating only compact intermediate representations (Hybrid Inference) .
Figure 2: Elastic architecture and Transition-Aware Triplet Training (T3). (a) Weight slicing. All width configurations share a single set of parameters: for a layer weight W∈Rwmax×wmax , a width- w subnetwork uses the top-left slice W(w)=W[1:w,1:w] , so all configurations are nested submatrices of the same W . (b) T3 training. At each iteration we sample a width pair (ws,wl) and a split layer k , and jointly optimize three paths over the same masked input: a small standalone encoder Eθ(ws) , a large standalone encoder Eθ(wl) , and a hybrid path that runs the first k layers at ws and transitions to wl at layer k via the elastic operator Pws→wl . Each encoder feeds a predictor Gθ together with a learned mask token m ; the three paths produc Lsmall , Lhybrid , and Llarge , all regressed against targets from a frozen V-JEPA teacher T on the unmasked video. The predictor width is set by a capped rule wG=min(wGmax,21wE) (Eq. 9 ) and is therefore not tied to the encoder width; the labels “Small”/“Large” denote which path a predictor serves, not its capacity.
Model
W
Hybrid
L
Params Range (M)
∣C∣
EV-JEPA-B
W1:{192,384,768,1024}
w→1024
12
5.6–152.6
7
EV-JEPA-L
W2:{192,384,768,1024,1280}
w→1280
12
5.6–238.0
9
EV-JEPA-H
W3:{192,384,768,1024,1280,1408}
w→1408
14
6.5–335.5
11
Table 1: Elastic Model Setup. In the default setting, every smaller width transitions to the largest width in its range, i.e., w→wmax,∀w∈W∖{wmax} . The total number of configurations is ∣C∣=∣W∣+(∣W∣−1)⋅K , where K is the number of candidate split layers. Reported values correspond to K=1 .
Student
Layers
wmax
Warm/Const/Decay
Peak LR
Clip
Batch/GPU
EV-JEPA-B/L
12
1024/1280
40 / 200 / 60
5.25×10−4
0.025
24
EV-JEPA-H
14
1408
40 / 200 / 60
5.25×10−4
0.025
24
CNN
20
384
20 / 80 / 200
3×10−3
0.025
12
Table 2: Optimization schedules by student family. All runs decay to a final learning rate of 5×10−5 .
Figure 3: SSv2 standalone and hybrid inference across increasing elastic ranges. Columns correspond to a) EV-JEPA-B ( W1 ), b) EV-JEPA-L ( W2 ), and c) EV-JEPA-H ( W3 ). The top row reports standalone elastic inference as a function of model parameters. The bottom row reports many-to-one hybrid execution w→wmax , where the first half of the encoder runs at source width w , and the remaining layers run at wmax . In the bottom row, T3 EV-JEPA w and T3 EV-JEPA wmax are shown as w -only and full-width T3 references, respectively, to measure hybrid recovery.
Figure 4: Standalone and hybrid inference on K400 for action recognition and EK100 for action anticipation. Standalone panels (1st, 3rd) report elastic inference as a function of model parameters. Hybrid panels (2nd, 4th) report hybrid execution w→wmax , where the first half of the encoder runs at source width w , and the remaining layers continue at wmax .
Figure 5: SSv2 any-to-any hybrid inference. In the any-to-any setting, a model executing at any source width ws can upscale at any larger width wl>ws . On the Hybrid Inference plots, each curve fixes a destination wl and sweeps the source ws . The star and dotted line mark the standalone accuracy of that destination width. Across all pairs, hybrid execution recovers close to each destination’s standalone accuracy, even at ws=192 , which is within ∼ 1 point of the standalone largest width. Accuracy is therefore governed by the destination (continuation) width, confirming that a small device can pair with any larger device at minimal loss.
Figure 6: Exit point, source width, and elastic token compression trade-offs. (left) SSv2 Top-1 for hybrid execution as a function of the exit layer k , for source widths w∈{192,384,768} and target width wl=1024 . (middle) SSv2 Top-1 versus k for the single path 192→1024 under different elastic compression ratios. (right) Transmission latency at each compression ratio over a theoretical 1 Mbps BLE link, computed as the transmitted payload divided by link bandwidth.
Figure 7: Elastic token compression for hybrid execution. A learned compressor is inserted at the exit point k of the source device. The ’compress module’ compresses the intermediate representation before transmission, reducing the number of tokens; the ’decompress module’ reconstructs the token grid on the target device, after which the shared elastic projection Pws→wl upscales to wl , and the remaining L−k layers execute. During training, all modules except the compress–decompress pair remain frozen.
τ
Kernel
Stride
Padding
Latent grid ( T′×H′×W′ ) / tokens
ρ
8×
(4,4,4)
(2,2,2)
(1,1,1)
4×8×8 / 256
64×
16×
(6,4,4)
(4,2,2)
(1,1,1)
2×8×8 / 128
128×
32×
(4,6,6)
(2,4,4)
(1,1,1)
4×4×4 / 64
256×
64×
(6,6,6)
(4,4,4)
(1,1,1)
2×4×4 / 32
512×
Table 3: Per-ratio configurations of the elastic encoder–decoder for an input grid of 8×16×16=2048 tokens. Kernel, stride, and padding are given as (time, height, width). τ=N/N′ is the token-reduction factor set by the sliced kernel (Eq. 15 ), and ρ is the resulting payload compression ratio relative to raw RGB (Eq. 18 ), which is the quantity reported in Fig. 6 .
Motion
Appearance
Method
Param.
SSv2
K400
Avg.
Image encoders Evaluated Using Same Protocol
DINOv2 [ 30 ]
1.1B
50.7
83.6
67.2
PE core G [ 3 ]
1.9B
55.4
88.5
72.0
SigLIP2 [ 34 ]
1.2B
49.9
87.3
68.6
Video encoders Evaluated Using Same Protocol
Table 4: Comparison with state-of-the-art models for action recognition on SSv2 and K400. Bold marks the best result per column, underline the second best. Avg. is the mean of SSv2 and K400.
Action anticipation (EK100)
Method
Param.
Verb
Noun
Action
InAViT [ 31 ]
160M
51.9
52.0
25.8
Video-LLaMA [ 41 ]
7.0B
52.9
52.0
26.0
PlausiVL [ 29 ]
8.0B
55.6
54.2
27.6
Frozen backbone
V-JEPA 2 ViT-L
300M
57.8
53.8
32.7
Table 5: Action anticipation on EK100. Bold marks the best result per column. Frozen-backbone rows attach an anticipation head to a frozen encoder.
Vision Language Action (VLA) models are mainstream in embodied intelligence but face high inference costs. Edge-Cloud Collaborative (ECC) inference offers an effective fix by easing edge-device computing pressure to meet real-time needs. However, existing ECC frameworks are suboptimal for VLA models due to two challenges: (1) Mainstream environment-oriented edge-cloud partitioning methods are susceptible to interference from visual noise; (2) Existing edge-cloud partitioning methods overlook the step-wise redundancy unique to embodied tasks, thereby disrupting the physical continuity of motion. To address these issues, we propose a novel ECC inference framework, termed RAPID. Specifically, we developed an implementation tailored to the proposed framework. Experiments demonstrate this achieves a speedup of up to 1.73x with only 5%~7% overhead.
Zihao Zheng, Sicheng Tian, Hangyu Cao +7
School of Computer Science, Peking University, Beijing, China. · School of Artificial Intelligence, Beijing Normal University, Beijing, China. · School of Software Engineering, South China University of Technology, Guangzhou, China. +1
Vision-Language-Action (VLA) models are mainstream in embodied intelligence but face high inference costs. Edge-Cloud Collaborative (ECC) deployment offers an effective fix by easing edge-device computing pressure to meet real-time needs. However, existing ECC frameworks are suboptimal for VLA models due to two challenges: (1) Diverse model structures hinder optimal ECC segmentation point identification; (2) Even if the optimal split point is determined, changes in network bandwidth can cause performance drift. To address these issues, we propose a novel ECC deployment framework for various VLA models, termed RoboECC. Specifically, we propose a model-hardware co-aware segmentation strategy to help find the optimal segmentation point for various VLA models. Moreover, we propose a network-aware deployment adjustment approach to adapt to the network fluctuations for maintaining optimal performance. Experiments demonstrate that RoboECC achieves a speedup of up to 3.28x with only 2.55%~2.62% overhead.
Zihao Zheng, Hangyu Cao, Jiayu Chen +6
School of Computer Science, Peking University, Beijing, China · School of Computer Science, South China University of Technology, Guangzhou, China · School of Artificial Intelligence, Beijing Normal University, Beijing, China +1
Training large language models (LLMs) is resource-intensive, and adapting them for diverse deployment scenarios with varying computational constraints remains challenging. While elastic architectures enable flexible model deployment and sparsely activated models allow input-adaptive computation, existing approaches treat these dimensions independently. Moreover, models catered towards on-device edge inference need to conform to the memory and compute limitations of the serving devices. In this paper, we introduce a unified framework that combines elastic structures with sparsely gated architectures to create models that adapt simultaneously to both deployment constraints and task requirements. Our approach employs a model backbone that conditions on both the context and target efficiency specifications, enabling fine-grained control over the accuracy-efficiency trade-off at inference time. The model learns to activate task-relevant parameters within elastically-nested sub-networks, allowing a single model to span multiple capacity points while maintaining input-adaptive routing. Through experiments we demonstrate that we can create a model that allows the flexibility to use 1,2,3,4 billion parameters while being more accurate than their dense counter-parts (2-5% on knowledge-intensive benchmarks) and at par with their static versions while delivering similar latency metrics as dense models. Overall, we save on device disk space by sharing the model parameters, allow flexibility of serving based on DRAM and compute available while delivering more accurate results.