JEPA · 2026-10-14

EV-JEPA: Elastic V-JEPA for Any Device and Cross-Device Inference

Anxhelo Diko, Jieyu Lin, Stefano Ambrogio, Ziyun Li, Mahmoud Assran, Th\'eo Moutakanni, Barbara De Salvo, Chiao Liu

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

EV-JEPA trains a single elastic V-JEPA model once for both standalone width changes and hybrid cross-device execution. Using Transition-Aware Triplet Training (T3), it jointly optimizes a smaller standalone path, a larger standalone path, and a hybrid path that switches width at a split layer where device handoff occurs. The same weights support many-to-one and variable-split configurations across benchmarks.

Why it matters

The authors aim to cover both width elasticity and collaborative cross-device inference in one model, addressing limitations where prior methods support only one axis. They report that EV-JEPA stays within 1 accuracy point of independently trained standalone baselines, improves over prior elastic training by up to 11 points, and recovers to within 1.1 points of the highest-capacity standalone model under narrow-to-wide cross-device hybrid execution.

Method

Elastic V-JEPA architecture with shared weights via nested parameterization (weight slicing) so width can be reconfigured at inference time and computation can be split across devices via an elastic transition projection P.
Transition-Aware Triplet Training (T3): each iteration samples (w_s, w_l) and a split layer k, and optimizes three paths—small standalone, hybrid (early layers narrow then transition), and large standalone—so cross-width compatibility is trained directly.
V-JEPA-style masked latent prediction with a frozen pretrained V-JEPA 2 encoder teacher; student regresses masked-token representations using an ℓ1 regression loss across the three T3 configurations; predictor width follows w_G=min(w_G^max, 1/2 w_E).

Limitation

EV-JEPA is tied to a teacher because training requires a pretrained V-JEPA model; the stability of a frozen teacher across widths without an external teacher (e.g., via an EMA scheme adapted to the elastic regime) is left to future work.

Abstract (from arXiv)

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.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-09 · Indexed by Pub-AI: 2026-10-14 UTC

preprint