Artificial Intelligence · arXiv:2608.15502 v1 · submitted 16 Aug 2026 · 24 references
EcoVLA: Energy-Efficient Device-Edge Co-Inference for Vision-Language-Action Models under Real-Time Constraints
The paper says
“Vision-Language-Action (VLA) models have emerged as a promising foundation for Embodied AI, but their high inference cost poses significant challenges for deployment in robotic systems.”
more of the abstract
In practice, on-device inference is constrained by limited compute capacity and energy budgets, struggling to simultaneously satisfy real-time control and energy efficiency requirements. Alternatively, offloading the inference workload to an edge server is susceptible to fluctuations in system conditions, introducing unpredictable latency risks. Device-edge co-inference offers a promising solution, but systematic research tailored to VLA models remains scarce, particularly a unified co-inference framework that jointly addresses real-time constraints and system-level energy efficiency. Thus, we propose EcoVLA, an adaptive device-edge co-inference framework for VLA models that maximizes system energy efficiency under real-time constraints. EcoVLA first introduces a unified stage-level abstraction over different VLA paradigms, establishing an architecture-agnostic co-inference design space. It then formulates a joint device-edge-network latency and energy prediction model to enable rapid runtime evaluation of candidate co-inference schemes. Building on this, EcoVLA continuously selects the energy-optimal scheme satisfying real-time constraints with millisecond-level overhead, adapting to runtime variations in network and system states. Furthermore, EcoVLA incorporates a lightweight transmission mechanism for inter-stage intermediate tensors to reduce the communication overhead incurred by cross-device collaboration. Experimental results across VLA models show that EcoVLA improves system energy efficiency by up to 236% over existing co-inference approaches under a 20 Hz action output frequency constraint, while consistently maintaining SLO satisfaction under dynamic network and edge workload conditions.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 27 Aug 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. EcoVLA is a serious VLA co-inference systems paper with a useful stage-level abstraction, but its formal objective omits the idle energy its own introduction says should count, so the headline efficiency gains are measured against the wrong metric. This line is wrong
3 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 24 references say.
The paper says, at Sec. 1 and Sec. 6.1 (Eqs. 3–5)
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The introduction states that system-level energy efficiency must be evaluated over the entire control cycle, including structural idle time, but the optimization objective in Eq. (5) maximizes L_act(pi)/E(pi) with E(pi) from Eq. (3) summing only per-stage compute energy and cross-device communication energy.
more of objection 1
No term accounts for idle or standby power during T_SLO - T(pi). The paper's own pi0.5 example (EcoVLA-fast 712 ms/8.15 J vs EcoVLA-opt 902 ms/6.0 J) shows that feasible plans differ in latency, so maximizing active-energy efficiency is not equivalent to maximizing cycle-level Actions/J unless idle power is zero. The central claim of maximizing system-level energy efficiency is therefore not supported by the formalization or by the measurements, which integrate power only over inference windows (Sec. 7.1). The authors should extend Eq. (3) with an idle-energy term, re-derive the scheduling rule, and re-run the evaluations with idle power included.
The paper says, at Sec. 7.2
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The reported gains (58% to 236%) and the claim of consistently maintaining the SLO are presented without error bars or multiple-trial statistics; each value appears to be a single measurement. Furthermore, the only co-inference baseline is a single Fixed Split from VLA-Perf [4]; no comparison is made to other adaptive co-inference or runtime-adaptation methods.
more of objection 2
Given that the central contribution is an empirical efficiency claim, the absence of variance information and the limited baseline set make the quantitative headline ('up to 236%') difficult to assess.
The paper says, at Sec. 5.3 and Sec. 7.2
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The prediction-accuracy evaluation (MAPE 6% latency, 3.8% energy) is conducted on the same platform used for offline profiling and online EWMA calibration. This validates the model on the calibration platform but does not establish the claimed architecture-agnostic transferability; stage profiles, residual terms, and queuing parameters may be platform-specific. The paper should at least reason about or measure sensitivity to platform changes, or soften the architecture-agnostic claim.
4 minor objections
The paper says, at General
Pith says: minor objection 1 of 4
The manuscript does not state whether code or data will be released; given the empirical nature of the central claims, a reproducibility statement or artifact URL would strengthen the paper.
The paper says, at Fig. 3
Pith says: minor objection 2 of 4
The caption reports Actions/J but does not indicate how many repeated runs the plotted values represent; the figure would benefit from error bars or a statement that they are omitted for clarity.
show 2 more
The paper says, at Fig. 4
Pith says: minor objection 3 of 4
The axes and legend are not fully self-contained in the caption; please state the unit of the horizontal axis (time steps) and define the shaded regions, if any.
The paper says, at Sec. 7.2
Pith says: minor objection 4 of 4
The paper should clarify which split point is used for the Fixed Split baseline from VLA-Perf [4] and whether that split was tuned for the testbed or taken as a default.
What would settle it
Measure the steady idle power P_idle of the full AGX Orin + RTX 4090 system over a 1 s control period with no inference. For every feasible mapping in the π0.5 and OpenVLA experiments, recompute E_tot(π)=E(π)+P_idle·(T_SLO−T(π)) using the reported or measured E and T values, then rerun the Eq. (5) argmax. If the selected plan changes for any model or load condition, or if the reported Actions/J ranking changes, the headline efficiency gains must be recomputed; if the argmax is invariant for P_idle swept from 0 to twice the measured value, the concern is settled.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 27 Aug 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 4 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
I read EcoVLA. It's worth a look, but there's a load-bearing inconsistency in the objective. The intro argues that system-level energy efficiency in robot control loops should be evaluated over the entire control cycle, because inference that finishes early leaves structural idle time. But Eq. (3)–(5) formalize efficiency as L_act(pi)/E(pi), where E(pi) sums only per-stage compute energy and cross-device communication energy. There is no idle/standby term. So the scheduler maximizes actions per inference-joule, not per control-cycle joule. Their own pi0.5 example shows the problem: EcoVLA-fast at 712 ms / 8.15 J and EcoVLA-opt at 902 ms / 6.0 J. For a 1 s SLO, cycle-level energy includes idle power over the time each plan leaves unused. With an idle power above roughly 11 W, the fast plan is actually the more cycle-efficient choice. On a Jetson AGX Orin, idle power is plausibly in that range. So the 236% claim is evaluated against a metric that the paper itself says is wrong, and the scheduler may select a genuinely suboptimal plan in system-level terms.
That said, the paper has real substance. The stage-level DAG abstraction over heterogeneous VLA paradigms—autoregressive and diffusion—is a sensible way to open up a co-inference design space, and the runtime with online load/network sensing plus the communication middleware is coherent. The math is simple and internally consistent, and enumeration is fine for the design-space sizes considered. The evaluation covers five VLA models and shows gains over fixed splitting and edge-only offloading.
The other soft spots are reproduction-type concerns. No code or data are released; there are no error bars; the only co-inference baseline is a single fixed split; and the >95% prediction accuracy is measured on the same platform whose stage profiles were fitted, with the online residual in Eq. (2) doing calibration. That makes the prediction claim less impressive than it looks. The paper also underspecifies how edge-stage batching is scheduled and costed, so I couldn't verify the claimed millisecond-level overhead.
For peer review, yes—a serious editor should send this out. The idle-energy inconsistency is fixable: add idle power to the energy model or measure cycle-level energy directly, then re-run. If the gains survive, this is a solid contribution to VLA co-inference and edge-AI deployment. I'd cite it if I worked in that area, and I'd bring it to reading group to debate the energy metric.
The strongest objection
Sec. 1 motivates the framework by the 'structural idle time' in robot control loops and states that system-level energy efficiency must be evaluated over the entire control cycle, not from instantaneous power or raw inference latency. However, the optimization objective in Eq. (5) uses η(π)=L_act(π)/E(π) (Eq. 4), where E(π) (Eq. 3) is the sum of per-stage compute energy and cross-device communication energy only. No term accounts for device/edge idle power during the unused portion of the period, T_SLO−T(π). For a fixed control period, total cycle energy is E(π)+P_idle·(T_SLO−T(π)); minimizing E(π) is equivalent to maximizing cycle-level Actions/J only if P_idle=0 or all feasible plans have identical T(π). The paper's own π0.5 example (EcoVLA-fast: 712 ms/8.15 J, EcoVLA-opt: 902 ms/6.0 J) shows feasible plans differ in latency, so the equivalence does not hold. The scheduler may select a plan that is optimal in inference-active energy but not in system-level energy efficiency as defined in the introduction. Until the model and measurements include idle energy, the central claim is unverified.
I read the paper in good faith: it is a serious systems paper with a plausible stage-level abstraction, a real testbed, and a coherent runtime design. The formal scheduling problem and the execution engine are internally consistent. My concern is not about experimental honesty or lack of artifacts (though those remain valid reasons for conditionality), but about a mismatch between the stated objective and the formalized one. The introduction explicitly argues that system energy efficiency must be measured over the whole control cycle because of structural idle time, yet Eqs. (3)–(5) optimize only the energy consumed during inference and communication. For a fixed 20 Hz period, the missing term P_idle·(T_SLO−T(π)) cancels only if all candidate plans have the same latency or idle power is zero; the paper's own π0.5 comparison shows different latencies. A simple idle-power measurement and recomputation would settle whether the claimed 236% and the energy-optimal plan rankings survive. I therefore keep the reader's CONDITIONAL verdict, but would add this correction as an explicit acceptance condition. Secondary concerns (mean-vs-tail SLO validation, missing artifacts and error bars) reinforce conditionality but are not the primary issue.
The referee's report
If the claims hold, EcoVLA would be a practically useful contribution: it couples a paradigm-agnostic stage abstraction with runtime energy-aware scheduling for VLA co-inference, and it includes a real implementation with measurements on a plausible device-edge platform. The paper also makes sensible design choices, such as separating control and data planes and preserving stage boundaries for model predictiveness. However, the central metric mismatch between the stated system-level energy objective and the formalized optimization objective prevents those strengths from being fully credited at present.
Pith's reading of the paper (a machine's summary, not the authors')
EcoVLA argues that device-edge co-inference for vision-language-action (VLA) models should be treated not as a one-time model split but as a continuous runtime scheduling problem: for each request, choose the stage-to-device mapping that maximizes output actions per joule while keeping predicted end-to-end latency within a real-time budget. The paper builds this on a stage-level graph abstraction that covers both autoregressive and diffusion VLA architectures, a joint latency-energy prediction model that combines offline profiling with online network and edge-load sensing, and a low-overhead scheduler that re-evaluates and re-ranks candidate mappings as conditions change. On a real testbed pairing a low-power embedded robot module with a server GPU, the authors report up to 236% higher energy efficiency than fixed co-inference baselines at a target 20 Hz action output frequency, with reported prediction errors around 6% for latency and 3.8% for energy. The broader point is that for continuous robot control, actions per joule within a deadline, rather than raw throughput or peak latency, is the metric that decides how long and how cheaply a robot can act.
What carries the argument. The central mechanism is a stage-level VLA execution graph: each VLA inference is decoupled into stages that must be serializable, relocatable, and independently executable, with a structured intermediate-state packet carrying stage input, stage output, and incremental context on every edge. This abstraction lets one co-inference design space cover heterogeneous VLA paradigms, and it exposes the exact byte sizes that the communication model needs. On top of it, the paper builds a joint latency-energy predictor that recursively computes stage start and finish times from device availability, parent-stage dependencies, cross-device transfer time $\hat{\tau}_{wv}$, and queue states, then adds an online residual for calibration; predicted energy sums per-stage power-delay products plus communication energy. The scheduler enumerates candidate mappings from the feasible space, filters by the SLO constraint $\hat{T}(\pi) \le T_{\mathrm{SLO}}$, and selects the maximum predicted actions-per-joule, falling back to minimum-latency when nothing meets the deadline, with a schedule-preserving runtime and dedicated data-plane middleware to ensure the chosen plan actually runs as modeled.
Core claim. The paper's central claim, formalized in Eq. (5), is that the optimal co-inference plan for a VLA request is the feasible stage mapping $\pi$ that maximizes predicted energy efficiency $\eta(\pi)=L_{\mathrm{act}}(\pi)/\hat{E}(\pi)$ subject to the predicted latency $\hat{T}(\pi) \le T_{\mathrm{SLO}}$, and that such a plan can be selected online with millisecond-level overhead. To support this, the paper proposes that VLA computation graphs be decoupled into serializable, relocatable, independently executable stages connected by a unified intermediate-state interface, so that autoregressive and diffusion-based models are governed by the same mapping space. It further claims that the joint latency-energy model, combining offline stage profiling, live bandwidth probing, edge queue prediction, and an exponentially weighted residual correction, predicts end-to-end latency and energy with mean absolute percentage errors of 6% and 3.8% respectively, and that this accuracy is enough for the scheduler to pick SLO-feasible, energy-optimal plans. The experiments claim energy-efficiency gains over fixed co-inference of 58% to 236% across five VLA models at a 20 Hz constraint, while satisfying the SLO under injected network degradation and variable edge workload.
The premise it rests on. The load-bearing assumption is that a VLA inference can be cut into stages whose individual computation times, power draws, and cross-device packet sizes are stable enough that offline profiling plus a small online correction predicts total latency and energy well; if stage costs depend strongly on input content or serializing model state costs more than modeled, the chosen plan can miss the deadline or waste energy.
If the paper is right (Pith's inferences, not the authors')
Fixed device-edge splits for VLA models become unreliable under changing network or load conditions, so future deployments should treat co-inference as an online optimization with re-evaluation per request.
With a 20 Hz real-time budget, a device-only plan can look energy-efficient in isolation yet violate the deadline; energy-efficiency claims for robot models should be reported only for SLO-satisfying configurations.
The value of runtime adaptation grows as edge contention increases, since the reported latency and energy reductions over static partitioning are largest under heavy load.
Intermediate-tensor transmission cost is a first-class term in the objective, so protocol choices such as pre-allocated buffers and an independent data channel directly change the achievable actions per joule.
Because action chunk length is fixed by the control window, minimizing end-to-end energy per request is equivalent to maximizing actions per joule, making the energy model the decisive scheduling input.
Editor's inference: the same stage-level abstraction could be extended to multiple edge servers, turning the binary device-edge mapping space into a placement problem over a heterogeneous fleet.
Editor's inference: the reported prediction accuracy suggests the profiling-plus-residual model could double as an offline capacity-planning tool, answering which robot hardware and edge GPU sizing meet a target actions-per-joule before deployment.
Editor's inference: a direct testable extension would compare EcoVLA's online choices against an oracle that knows true future latency and energy, isolating how much of the 236% gain comes from model fidelity versus from the adaptation policy itself.
Editor's inference: the actions-per-joule framing implies benchmarking VLA deployment by task endurance, not just control frequency; for long-horizon manipulation, a slightly slower plan that uses far less energy may enable more total work within a battery budget.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Vision-Language-Action (VLA) models have emerged as a promising foundation for Embodied AI, but their high inference cost poses significant challenges for deployment in robotic systems. In practice, on-device inference is constrained by limited compute capacity and energy budgets, struggling to simultaneously satisfy real-time control and energy efficiency requirements. Alternatively, offloading the inference workload to an edge server is susceptible to fluctuations in system conditions, introducing unpredictable latency risks. Device-edge co-inference offers a promising solution, but systematic research tailored to VLA models remains scarce, particularly a unified co-inference framework that jointly addresses real-time constraints and system-level energy efficiency. Thus, we propose EcoVLA, an adaptive device-edge co-inference framework for VLA models that maximizes system energy efficiency under real-time constraints. EcoVLA first introduces a unified stage-level abstraction over different VLA paradigms, establishing an architecture-agnostic co-inference design space. It then formulates a joint device-edge-network latency and energy prediction model to enable rapid runtime evaluation of candidate co-inference schemes. Building on this, EcoVLA continuously selects the energy-optimal scheme satisfying real-time constraints with millisecond-level overhead, adapting to runtime variations in network and system states. Furthermore, EcoVLA incorporates a lightweight transmission mechanism for inter-stage intermediate tensors to reduce the communication overhead incurred by cross-device collaboration. Experimental results across VLA models show that EcoVLA improves system energy efficiency by up to 236% over existing co-inference approaches under a 20 Hz action output frequency constraint, while consistently maintaining SLO satisfaction under dynamic network and edge workload conditions.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 0.98
“Experimental results across VLA models show that EcoVLA improves system energy efficiency by up to 236% over existing co-inference approaches under a 20 Hz action output frequency constraint, while consistently maintaining SLO satisfaction under dynamic network and edge workload conditions.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.22 · states limitation 0.10 · human participants 0.03 · controlled comparison 0.79 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Vision-Language-Action (VLA) models have emerged as a promising foundation for Embodied AI, but their high inference cost poses significant challenges for deployment in robotic systems. | background 1.00 | 0.04 | 0.02 | asserted 0.88 |
| In practice, on-device inference is constrained by limited compute capacity and energy budgets, struggling to simultaneously satisfy real-time control and energy efficiency requirements. | background 1.00 | 0.07 | 0.02 | asserted 1.00 |
| Alternatively, offloading the inference workload to an edge server is susceptible to fluctuations in system conditions, introducing unpredictable latency risks. | background 0.98 | 0.06 | 0.01 | asserted 0.92 |
| Device-edge co-inference offers a promising solution, but systematic research tailored to VLA models remains scarce, particularly a unified co-inference framework that jointly addresses real-time constraints and system-level energy efficiency. | gap 1.00 | 0.06 | 0.05 | asserted 0.93 |
| Thus, we propose EcoVLA, an adaptive device-edge co-inference framework for VLA models that maximizes system energy efficiency under real-time constraints. | contribution 1.00 | 0.86 | 0.34 | asserted 0.99 |
| EcoVLA first introduces a unified stage-level abstraction over different VLA paradigms, establishing an architecture-agnostic co-inference design space. | method 0.72 | 0.92 | 0.04 | asserted 1.00 |
| It then formulates a joint device-edge-network latency and energy prediction model to enable rapid runtime evaluation of candidate co-inference schemes. | method 0.92 | 0.90 | 0.03 | asserted 0.98 |
| Building on this, EcoVLA continuously selects the energy-optimal scheme satisfying real-time constraints with millisecond-level overhead, adapting to runtime variations in network and system states. | method 0.94 | 0.91 | 0.16 | asserted 1.00 |
| Furthermore, EcoVLA incorporates a lightweight transmission mechanism for inter-stage intermediate tensors to reduce the communication overhead incurred by cross-device collaboration. | method 0.81 | 0.88 | 0.03 | asserted 1.00 |
| Experimental results across VLA models show that EcoVLA improves system energy efficiency by up to 236% over existing co-inference approaches under a 20 Hz action output frequency constraint, while consistently maintaining SLO satisfaction under dynamic network and edge workload conditions. | result 1.00 | 0.96 | 0.86 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 7,700 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (5)
Figures are the authors', reproduced from the paper.
References (24)
- [4] Jiang, W., et al.: How fast can I run my VLA? Demystifying VLA inference performance with VLA-Perf. arXiv preprint arXiv:2602.18397 (2026) on PitharXiv
- [1] Black, K., et al.:π 0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164 (2024) on PithPith's reviewarXiv
- [2] Black, K., et al.:π 0.5: A vision-language-action model with open-world generalization. Proceedings of Machine Learning Research (2025) on Pith
- [3] Dai, Y., et al.: ActionFlow: A pipelined action acceleration for vision language models on edge. arXiv preprint arXiv:2512.20276 (2025) on PithPith's reviewarXiv
- [5] Kang, Y., et al.: Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. In: Proceed- ings of ASPLOS (2017) on Pith
- [6] Kim, M.J., et al.: OpenVLA: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246 (2024) on PithPith's reviewarXiv
- [7] Li, Y., et al.: SP-VLA: A joint model scheduling and token pruning approach for VLA model acceleration. In: Proceedings of ICLR (2026) on Pith
- [8] Liu, S., et al.: RDT-1B: A diffusion foundation model for bimanual manipulation. In: Proceedings of ICLR (2025) on Pith
- [9] Ma, Y., et al.: Running VLAs at real-time speed. arXiv preprint arXiv:2510.26742 (2025) on PitharXiv
- [10] Ning, J., et al.: DSSD: Efficient edge-device LLM deployment and collaborative inference via distributed split speculative decoding. In: Proceedings of ICML (2025) on Pith
- [11] Qu, G., et al.: Mobile edge intelligence for large language models: A contemporary survey. IEEE Commu- nications Surveys & Tutorials (2025) on Pith
- [12] Sapkota, R., et al.: Vision-language-action models: Concepts, progress, applications and challenges. arXiv preprint arXiv:2505.04769 (2025) on PitharXiv
- [13] Shukor, M., et al.: SmolVLA: A vision-language-action model for affordable and efficient robotics. arXiv preprint (2025) on Pith
- [14] Tang, J., et al.: VLASH: Real-time VLAs via future-state-aware asynchronous inference. arXiv preprint arXiv:2512.01031 (2025) on PithPith's reviewarXiv
- [15] Team, O.M., et al.: Octo: An open-source generalist robot policy. In: Proceedings of RSS (2024) on Pith
- [16] Wen, J., et al.: TinyVLA: Towards fast, data-efficient vision-language-action models for robotic manipula- tion. IEEE RA-L (2025) on Pith
- [17] Xu, S., et al.: VLA-Cache: Towards efficient vision-language-action model via adaptive token caching in robotic manipulation. In: Proceedings of NeurIPS (2025) on Pith
- [18] Yang, F., et al.: Efficient deployment of large language model across cloud-device systems. In: Proceedings of SOCC (2024) on Pith
- [19] Yang, Y., et al.: EfficientVLA: Training-free acceleration and compression for vision-language-action mod- els. arXiv preprint arXiv:2506.10100 (2025) on PithPith's reviewarXiv
- [20] Yuan, C., et al.: Task-oriented feature compression for multimodal understanding via device-edge co- inference. IEEE TMC (2025) on Pith
show 4 more
- [21] Zhang, D., et al.: Pure vision language action (VLA) models: A comprehensive survey. arXiv preprint arXiv:2509.19012 (2025) on PitharXiv
- [22] Zhang, M., et al.: EdgeShard: Efficient LLM inference via collaborative edge computing. IEEE IoT Journal (2024) on Pith
- [23] Zhou, A., et al.: Graph neural networks automated design and deployment on device-edge co-inference systems. In: Proceedings of DAC (2024) on Pith
- [24] Zhou, A., et al.: ACE-GNN: Adaptive GNN co-inference with system-aware scheduling in dynamic edge environments. IEEE TCAD (2025) on Pith
How the literature uses it
What this paper does with the arXiv papers it cites
What it cites · 7 sentences without a confident label
Cites as support 1 paper
-
Even if inference completes early, the system must wait for the current action execution period before issuing the next command, a structural idle time also identified by VLASH [ref] .
VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference · Introduction · 1.00
Cites as background 8 papers
-
At the system level, ActionFlow [ref] and Real-time VLA [ref] lower on-device latency through operator fusion and pipelined scheduling, while VLASH [ref] overlaps inference with action execution to mitigate structural idle time.
ActionFlow: A Pipelined Action Acceleration for Vision Language Models on Edge · VLA Deployment and Acceleration · 0.90
-
At the system level, ActionFlow [ref] and Real-time VLA [ref] lower on-device latency through operator fusion and pipelined scheduling, while VLASH [ref] overlaps inference with action execution to mitigate structural idle time.
2510.26742 · VLA Deployment and Acceleration · 0.90
-
At the system level, ActionFlow [ref] and Real-time VLA [ref] lower on-device latency through operator fusion and pipelined scheduling, while VLASH [ref] overlaps inference with action execution to mitigate structural idle time.
VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference · VLA Deployment and Acceleration · 0.90
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment