Pith. sign in

REVIEW 2 major objections 5 minor 98 references

How Much Does It Cost to Answer My Question? Benchmarking Cloud VLM-based VQA Systems

T0 review · 2 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Client-side image preprocessing for cloud VQA is not intrinsically helpful: its effect on accuracy, latency, payload, and token cost depends on the target model, API paradigm, and task.

desk verdict Solid systems benchmark with real measurement work; the headline claim is probably right, but the run-count arithmetic doesn't add up and the accuracy deltas lack error bars. read the letter →

arxiv 2608.07861 v1 pith:GXCYWW2I submitted 2026-08-08 cs.CV cs.HCcs.IRcs.MM

Henri Vanhuynegem , Weitao Xu , Yiran Shen , Guohao Lan This is my paper · ORCID
classification cs.CVcs.HCcs.IRcs.MM
keywords cloud-basedVQAclient-sidepreprocessingtokenaccountinglatencymeasurementvision-languagemodelsgaze-guidedcroppingcost-qualitytrade-off
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

VQABench is a benchmark that treats client-side image preprocessing as the controlled variable in cloud-VLM-based visual question answering, asking what visual input a mobile assistant should offload when accuracy, latency, payload, and token cost all matter. Across 12 preprocessing techniques, 3 datasets, and 4 commercial VLMs (95,168 API calls), the paper finds that preprocessing effectiveness is not intrinsic to the technique: the same method can speed up one model–provider stack, barely affect another, and slow down a third. It also finds that smaller payloads do not mean fewer tokens, because providers bill image tokens on decoded pixel geometry rather than transmitted bytes, so compression-only methods save no tokens on pixel-based accounting. The practical upshot is that preprocessing strategy must be chosen jointly with model selection and task type, and that gaze-guided cropping without a thumbnail is the strongest cost-aware candidate on models whose token accounting rewards pixel reduction.

What carries the argument

The controlled comparison matrix is the central object: 12 preprocessing variants (baseline, five global-reduction methods, three region-selective methods, and three thumbnail-augmented variants) crossed with three datasets and four commercial VLMs, with each request measured on accuracy, payload, tokens, and decomposed latency. The mechanism that explains the results is provider token accounting on decoded pixel geometry rather than transmitted bytes: pixel-sensitive models (tile-based or area-proportional charges) reward cropping and downsampling, flat-rate per-image models charge the same for any single image and multiply charges for each additional image part, and at least one hybrid endpoint assigns images to two token buckets based on aspect ratio.

What would settle it

Conduct a human-rater study on the same DriVQA and VOILA-A outputs and compare per-technique rankings: if GazeROI no longer beats Downsampler or compression methods under human scores, the paper's accuracy conclusions for open-ended tasks fail; separately, submit the same source image at identical resolution encoded at JPEG quality 1 and quality 100 and check that provider-reported image tokens are identical, which would confirm the decoded-pixel-geometry accounting mechanism.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that client-side preprocessing reshapes the full cost-quality trade-off of a cloud-VLM VQA request across four dimensions—answer accuracy, payload size, token usage, and latency—and that the direction and size of the effect are governed by the target model, the API paradigm, the provider's token-accounting rule, and the task formulation. The evidence shows three regimes: compression-only methods preserve accuracy and shrink payload but save no tokens; cropping and downsampling save tokens only on providers that account per decoded pixel, while multi-image submissions (thumbnail-augmented or multi-crop) inflate tokens on fixed-rate and hybrid providers; and accuracy loss tracks what information a method discards—color and spatial detail cost less than removing the task-relevant region, with saliency-based and thumbnail-only inputs causing the largest drops and more hallucinated claims. A poorly chosen strategy can simultaneously raise cost and latency while lowering accuracy, and the paper recommends gaze-guided cropping as the knee point on the Pareto frontier for pixel-sensitive models.

Load-bearing premise

The load-bearing assumption is that the LLM-as-judge rubric scores for DriVQA and VOILA-A reflect genuine answer quality; if the judge mis-ranks answers on these open-ended tasks (the rubric weights are hand-set and no human-agreement check is reported), the accuracy-tier conclusions weaken even though the latency and token measurements stand.

Editorial extensions

If this is right

  • Developers should co-select preprocessing with model and provider rather than applying a fixed reduction recipe, because the same technique moves along different accuracy-latency-cost frontiers on different stacks.
  • Lossy compression at quality 85 is the safe default: it preserves near-baseline accuracy across models and cuts payload, while saving no tokens on pixel-accounting providers.
  • Multi-image methods such as thumbnail-augmented or multi-crop ROI submissions should be avoided on flat-rate and hybrid providers, where they can inflate token usage by up to 416%.
  • Gaze-guided cropping without a thumbnail is the recommended knee point on pixel-sensitive models, roughly halving tokens while holding accuracy close to baseline.
  • On fixed-budget providers such as the tested flash-tier model, the unprocessed baseline can dominate every preprocessing variant, so preprocessing should be skipped.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the aspect-ratio token-bucket finding implies a content-independent optimization—padding any non-square image to square before submission can cut image tokens by about 40% on that hybrid endpoint, a lever the benchmark's main analysis does not exploit.
  • Editorial inference: since provider token-accounting rules are undocumented and may change, the quantitative rankings are snapshots; the durable insight is the mechanism (decoded-pixel geometry vs. per-image charges), which should be re-tested when endpoints update.
  • Editorial inference: the accuracy reversals across models suggest a future system could route each user query to the model–preprocessing pair that maximizes expected quality per unit cost, treating preprocessing selection as part of model routing rather than a separate step.
  • Editorial inference: the saliency-crop failures imply that attention signals tied to user intent (gaze) are more reliable than bottom-up visual saliency for deciding what to keep, which suggests intent-conditioned cropping as a research direction.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper introduces VQABench, a benchmark that treats client-side image preprocessing as a controlled variable for cloud-VLM-based VQA. It evaluates 12 preprocessing techniques (compression, downsampling, grayscale, and gaze-, saliency-, and YOLO-based ROI methods with and without thumbnails) on three datasets (VQA-MHUG, DriVQA, VOILA-A) across four commercial VLMs from three providers, claiming 95,168 API calls. The central claim is that preprocessing effectiveness is not intrinsic to a technique but depends on the model–provider stack, API paradigm, provider token-accounting rule, and task formulation. The paper reports latency decompositions, token-usage breakdowns, answer-quality metrics, Pareto analyses, and scenario-specific deployment rankings, plus controlled token-accounting experiments in the appendices.

Significance. If the empirical results hold, the paper fills a real gap: prior VQA benchmarks focus almost exclusively on accuracy, and no existing benchmark treats preprocessing as a controlled variable across commercial cloud APIs. The controlled token-accounting experiments in Appendices H–J, which decouple decoded pixel geometry from transmitted payload size, are a strong and reusable contribution, and the public code and measurement artifacts support reproducibility. The observed provider-specific accounting rules (e.g., Gemini's flat per-image budget, Realtime-mini's square/non-square token buckets) are concrete, falsifiable findings of practical use. However, the central quantitative claims are currently weakened by an unexplained discrepancy in the reported API-call count, missing per-condition sample sizes, and the absence of confidence intervals or significance tests for the accuracy and latency deltas that support the model-dependence conclusion.

major comments (2)
  1. [Section VI.B, Section VI.C, Eq. (2)] The reported total of 95,168 API calls is inconsistent with the stated full-dataset protocol. The abstract and Section III state that 12 preprocessing techniques are evaluated on VQA-MHUG (N=3,990), DriVQA (N=24), and VOILA-A (N=500) across four VLMs. Under that protocol, VQA-MHUG alone requires 12 × 4 × 3,990 = 191,520 calls, already more than double the reported total; even if DriVQA and VOILA-A used only one model, at least 12 × 4 × 1,852 ≈ 88,896 calls would be needed for VQA-MHUG, implying a subsample of about 1,850 pairs rather than the stated 3,990. No subsampling is disclosed anywhere in the manuscript. Since Tables II–V report means without per-cell sample sizes, confidence intervals, or significance tests, the accuracy deltas that drive the central claim (e.g., Downsampler −6.6% on Realtime-mini vs. −23.4% on GPT-5.4 in Table V, or the latency ranking in Figure 6) cannot be verified against sampling error. Please disclose the exact per-condition sample sizes, clarify how the 95,168 total is obtained, and report confidence intervals or significance tests for the main accuracy and latency comparisons.
  2. [Section VI.B, Section VI.C, Eq. (2)] The open-ended quality conclusions for DriVQA and VOILA-A rest entirely on a GPT-4.1 LLM judge with hand-set rubric weights (0.50/0.25/0.15/0.10 for DriVQA in Section III.D and 0.40/0.30/0.20/0.10 in Eq. (2) for VOILA-A), yet the paper reports no human-agreement validation, no judge-stability analysis, and no alternative-judge sensitivity check. With only 24 DriVQA items, the observed differences (e.g., baseline 0.52 vs. GazeROI 0.50 in Figure 9) are within a plausible range of judge or sampling noise, and a different rubricing could change which techniques are considered to preserve quality. Because these results are used to claim that preprocessing effects depend on task formulation, please add a human-validation study or, at minimum, a judge-variance/sensitivity analysis and report uncertainties around the per-technique open-ended scores.
minor comments (5)
  1. [Table III] The header 'Reqest→ First response' contains a typo and should read 'Request→ First response'; the same table also uses inconsistent capitalization for the latency components.
  2. [Section III.B] The statement that Base64/data-URL wrapping inflates the payload by 'roughly 33%' is correct for pure binary-to-Base64 expansion but may be slightly different after JSON escaping and provider headers; a brief definition or formula would help reproducibility.
  3. [Section IV.A, Appendix D] Appendix D reports that identical client-side preprocessing techniques have 2–3× higher local preprocessing latency under the Realtime WebSocket API than under REST (Table XI). Since preprocessing is described as running locally before the request is sent, it is unclear how the API interface could affect it; please clarify whether the measured interval includes any connection/session overhead or whether the code path differs between the two interfaces.
  4. [Figure 4] The stacked accuracy distributions in Figure 4 are informative, but the 'Weighted Accuracy' labels on the right are not defined in the caption; please add a short explanation or a table with the same values.
  5. [Section VII.A, Figures 11 and 20-22] The Pareto analysis uses a normalized per-query token cost as the third objective, but the normalization formula is only described in the figure caption; please state it in the main text so the trade-off plots are reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are direct empirical measurements against external provider APIs and declared user-defined weights, not derivations that reduce to their own inputs.

full rationale

This paper is an empirical measurement study, not a derivation. The central claim that preprocessing effectiveness depends on the model-provider stack, API paradigm, token-accounting rule, and task formulation is supported by direct measurements of provider-reported latency, token usage, payload size, and answer-quality metrics under controlled preprocessing conditions. No parameter is fitted to a subset of data and then reused to predict a closely related quantity. The model-selection screening on 100 VQA-MHUG baseline samples chooses four models by baseline latency and accuracy; it does not fit preprocessing effects, and the final accuracy and latency results are separately measured on the full datasets. The LLM-as-judge rubric weights in Section III-D and the deployment-scenario weights in Table VI are declared user-defined inputs, not hidden fits used to manufacture predictions. Token-accounting explanations are induced from controlled experiments in Appendices H-J and then used to interpret, not to generate, the main token measurements. No load-bearing uniqueness theorem or ansatz is imported from prior work by these authors; citations to gaze-guided preprocessing and VOILA-A are external support for technique design, not for the benchmark conclusions. The discrepancy noted between the reported 95,168 API calls and the naive full-cross-product count is a data-integrity and sample-size concern, not a circularity, and does not enter this score. Therefore no circular step was found.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The benchmark's conclusions rest primarily on three unverified but reasonable assumptions: provider-reported telemetry is trustworthy, the accuracy metrics (including LLM-as-judge with hand-chosen weights) measure true answer quality, and the controlled campus-network pipeline represents mobile deployment. The scenario weights and judge weights are explicit user-defined inputs, not hidden fitted parameters. No invented physical or formal entities are introduced.

free parameters (3)
  • DriVQA LLM-judge weights = weighted match 0.50, majority alignment 0.25, ambiguity handling 0.15, contradiction rate 0.10
    Hand-chosen combination weights in Section III.D; no sensitivity analysis is reported, and they directly determine which preprocessing methods appear best on DriVQA.
  • VOILA-A LLM-judge weights = evidence consistency 0.40, coverage 0.30, detail quality 0.20, unsupported claims 0.10
    Hand-chosen weights in Eq. (2); no validation against human ratings is reported, yet these scores drive the accuracy-tier conclusions in Section VI.C.
  • Deployment scenario weight vectors = Seven (accuracy/latency/cost) triples from 0.60/0.10/0.30 to 0.20/0.40/0.40
    Hand-picked scenario priorities in Table VI; they are explicit user-defined inputs, but the per-scenario rankings and the headline recommendation that GazeROI ranks first in most scenarios are conditional on these weights.
assumptions (3)
  • domain assumption Provider-reported token counts and client-observed latency components are accurate, deterministic, and stable over the measurement period.
    Token-usage tables (Table IV) and latency decompositions (Tables II and III) treat provider-reported usage records and client timings as ground truth. The paper itself notes in Appendix A that commercial APIs may evolve, so the measurements are time-bounded snapshots.
  • domain assumption The three datasets' extraction and the LLM-as-judge rubrics are faithful proxies for answer quality.
    Section III-D defines accuracy via VQA soft accuracy and two GPT-4.1 judge rubrics. The rubric weights are hand-set and no human-agreement validation is reported, so the accuracy tiers could shift under a different judge or human rating protocol.
  • domain assumption The tested client pipeline, Base64 data URLs over a campus 1 Gbps Ethernet link, is representative of real mobile offloading.
    Section III-A and Appendix A describe this pipeline and network. Real mobile links add variance; the Wi-Fi/4G experiment in Appendix G covers only the baseline pipeline, so preprocessing interaction with realistic mobile networks is not measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Much Does It Cost to Answer My Question? Benchmarking Cloud VLM-based VQA Systems." pith.science (2026). https://pith.science/paper/GXCYWW2I

@misc{pith2026260807861,
  author       = {Pith},
  title        = {Pith review of: How Much Does It Cost to Answer My Question? Benchmarking Cloud VLM-based VQA Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GXCYWW2I}},
  note         = {Machine review of arXiv:2608.07861}
}
read the original abstract

Vision-language models (VLMs) are becoming a practical backend for mobile visual question answering (VQA) systems, enabling smartphones and smart glasses to answer users' questions about the physical world. Since modern VLMs remain difficult to run on mobile and edge devices, VQA systems increasingly offload inference to cloud-based VLMs. This gives mobile devices access to stronger computation, but it also makes visual input preparation a key system variable: how the image is prepared before offloading affects not only answer quality but also payload size, token cost, and system latency. Proprietary APIs expose little control over model internals or serving behavior, leaving client-side preprocessing as the main practical optimization space for downstream developers. Many such techniques have been proposed for visual offloading, yet their cost-quality impact on commercial cloud VLMs has never been studied. To fill this gap, we present VQABench, the first systematic benchmark that treats client-side input preprocessing as a controlled variable for cloud-VLM-based VQA. We evaluate 12 preprocessing techniques across three VQA datasets and four commercial VLMs from three providers, totaling 95,168 API calls. Our results show that preprocessing is not universally beneficial: its effectiveness depends on the target model, API paradigm, provider token-accounting rule, and task formulation. A poorly selected preprocessing strategy can increase deployment cost or latency while degrading answer accuracy. Overall, our benchmark clarifies when preprocessing helps, when it fails, and why, providing insights to guide future research and real-world deployment of VQA systems.

Figures

Figures reproduced from arXiv: 2608.07861 by the authors.

Figure 1
Figure 1. Illustration of proprietary VLM-based VQA. The user [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the VQA request pipeline. Four quantities are measured (orange): the payload size, the token usage, the answer accuracy, and two latency spans. B. The Cost Dimensions and Metrics In the end-to-end pipeline, we measure four quantities, highlighted in orange in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the preprocessing techniques applied [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: VQA soft-accuracy distribution for the 13 screened [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Latency–accuracy trade-off across 13 screened models. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: End-to-end latency for selected preprocessing tech [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Percentage change in provider-reported token usage [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Answer quality scores by preprocessing technique on [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: LLM-as-judge answer quality scores by preprocessing [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 12
Figure 12. Figure 12: Preprocessor rankings across deployment scenarios Columns ordered accuracy-priority latency/cost-priority; sub-labels are accuracy/latency/cost weights (×100). for Realtime-mini. Each cell reports a method’s rank under one scenario-specific weighting. substantial degr…
Figure 13
Figure 13. Figure 13: Representative image–question–answer triples from the three datasets, illustrating their different [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Mean serialized payload size compared with mean [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Mean serialized payload size compared with mean [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]
Figure 17
Figure 17. Figure 17: Latency by network connection. ing should be evaluated not only by median latency, but also by whether upper-percentile latency remains bounded. GazeROI is the strongest candidate under this criterion: it achieves one of the lowest median latencies and the lowest p99 …
Figure 18
Figure 18. Figure 18: Impact of image resolution (pixel count) on input [PITH_FULL_IMAGE:figures/full_fig_p020_18.png]
Figure 19
Figure 19. Figure 19: Impact of JPEG quality on input token and payload [PITH_FULL_IMAGE:figures/full_fig_p020_19.png]
Figure 12
Figure 12. Figure 12: Each cell shows a method’s rank (1 = best) within a scenario, columns run from accuracy-priority to latency/cost￾priority, and rows are ordered by mean rank. For GPT-5.4, the result is even stronger than for Realtime￾mini. GazeROI ranks first in all seven scenarios ( …
Figure 24
Figure 24. Figure 24: Preprocessor rankings across deployment scenarios Columns ordered accuracy-priority latency/cost-priority; sub-labels are accuracy/latency/cost weights (×100). for the Claude-Sonnet-4.6. (rank 2 in the five accuracy-to-balanced scenarios, rank 3 in the two most cost-d…
Figure 23
Figure 23. Figure 23: Preprocessor rankings across deployment scenarios Columns ordered accuracy-priority latency/cost-priority; sub-labels are accuracy/latency/cost weights (×100). for the GPT-5.4. Quality on budget 60/10/30 Accuracy first 50/25/25 Realtime quality 40/40/20 Balanced 33/33…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

98 extracted references · 69 canonical work pages

  1. [1]

    VQA: Visual Question Answering,

    S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh, “VQA: Visual Question Answering,” inInternational Conference on Computer Vision (ICCV), 2015

    2015

  2. [2]

    Making the V in VQA matter: Elevating the role of image understanding in visual question answering,

    Y . Goyal, T. Khot, D. Summers-Stay, D. Batra, and D. Parikh, “Making the V in VQA matter: Elevating the role of image understanding in visual question answering,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6904–6913, 2017

    2017

  3. [3]

    Project Aria: A new tool for egocentric multi-modal AI research,

    J. Engel, K. Somasundaram, M. Goesele, A. Sun, A. Gamino, A. Turner, A. Talattof, A. Yuan, B. Souti, B. Meredith, C. Peng, C. Sweeney, C. Wilson, D. Barnes, D. DeTone, D. Caruso, D. Valleroy,et al., “Project Aria: A new tool for egocentric multi-modal AI research,”arXiv preprint arXiv:2308.13561, 2023

    arXiv 2023

  4. [4]

    Ray-Ban Meta AI glasses gen 2 & gen 1

    Meta and Ray-Ban, “Ray-Ban Meta AI glasses gen 2 & gen 1.” Product page, 2025. Accessed 2026-05-11

    2025

  5. [5]

    Vision-language models for vision tasks: A survey,

    J. Zhang, J. Huang, S. Jin, and S. Lu, “Vision-language models for vision tasks: A survey,” 2024

    2024

  6. [6]

    A survey on multimodal large language models,

    S. Yin, C. Fu, S. Zhao, K. Li, X. Sun, T. Xu, and E. Chen, “A survey on multimodal large language models,”National Science Review, vol. 11, no. 12, p. nwae403, 2024. Review article

    2024

  7. [7]

    VizWiz grand challenge: Answering visual questions from blind people,

    D. Gurari, Q. Li, A. J. Stangl, A. Guo, C. Lin, K. Grauman, J. Luo, and J. P. Bigham, “VizWiz grand challenge: Answering visual questions from blind people,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3608–3617, 2018

    2018

  8. [8]

    Be My Eyes: Lend your eyes to the blind

    Be My Eyes, “Be My Eyes: Lend your eyes to the blind.” Mobile application, 2015

    2015

Show all 98 references
  1. [9]

    Long-form answers to visual questions from blind and low vision people,

    M. Huh, F. Xu, Y .-H. Peng, C. Chen, H. Murugu, D. Gurari, E. Choi, and A. Pavel, “Long-form answers to visual questions from blind and low vision people,” inConference on Language Modeling (COLM), 2024

    2024

  2. [10]

    Augmented reality anatomy visualization for surgery assistance with HoloLens,

    E. Castelan, M. Vinnikov, and X. Zhou, “Augmented reality anatomy visualization for surgery assistance with HoloLens,” inProceedings of the ACM International Conference on Interactive Media Experiences (IMX), pp. 329–331, 2021

    2021

  3. [11]

    LLMs enable context-aware augmented reality in surgical navigation,

    H. Javaheri, O. Ghamarnejad, P. Lukowicz, G. A. Stavrou,et al., “LLMs enable context-aware augmented reality in surgical navigation,”arXiv preprint arXiv:2412.16597, 2024

    2024 arXiv

  4. [12]

    DriVQA: A gaze- based dataset for visual question answering in driving scenarios,

    K. Rekanar, J. M. Joyce, M. Hayes, and C. Eising, “DriVQA: A gaze- based dataset for visual question answering in driving scenarios,”Data in Brief, vol. 59, p. 111367, 2025

    2025

  5. [13]

    Mimicking human attention in driving scenarios for enhanced visual question answering: Insights from eye-tracking and the human attention filter,

    K. Rekanar, M. J. Hayes, and C. Eising, “Mimicking human attention in driving scenarios for enhanced visual question answering: Insights from eye-tracking and the human attention filter,”Machine Learning with Applications, 2025. 14

    2025

  6. [14]

    Demystifying small language models for edge deployment,

    Z. Lu, X. Li, D. Cai, R. Yi, F. Liu, W. Liu, J. Luan, X. Zhang, N. D. Lane, and M. Xu, “Demystifying small language models for edge deployment,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), (Vienna, Austria)...

    2025

  7. [15]

    Efficient processing of deep neural networks: A tutorial and survey,

    V . Sze, Y .-H. Chen, T.-J. Yang, and J. S. Emer, “Efficient processing of deep neural networks: A tutorial and survey,”Proceedings of the IEEE, vol. 105, no. 12, pp. 2295–2329, 2017

    2017

  8. [16]

    LLM in a flash: Efficient large language model inference with limited memory,

    K. Alizadeh, S. I. Mirzadeh, D. Belenko, S. K. Khatamifard, M. Cho, C. C. Del Mundo, M. Rastegari, and M. Farajtabar, “LLM in a flash: Efficient large language model inference with limited memory,” in Proceedings of the 62nd Annual Meeting of the Association for Com- putationa...

    2024

  9. [17]

    Mo- bileLLM: Optimizing sub-billion parameter language models for on- device use cases,

    Z. Liu, C. Zhao, F. Iandola, C. Lai, Y . Tian, I. Fedorov, Y . Xiong, E. Chang, Y . Shi, R. Krishnamoorthi, L. Lai, and V . Chandra, “Mo- bileLLM: Optimizing sub-billion parameter language models for on- device use cases,” inProceedings of the 41st International Conference on ...

    2024

  10. [18]

    MobileVLM: A fast, strong and open vision language assistant for mobile devices,

    X. Chu, L. Qiao, X. Lin, S. Xu, Y . Yang, Y . Hu, F. Wei, X. Zhang, B. Zhang, X. Wei, and C. Shen, “MobileVLM: A fast, strong and open vision language assistant for mobile devices,”arXiv preprint arXiv:2312.16886, 2023

    2023 arXiv

  11. [19]

    Bench- marking tinyml systems: Challenges and direction,

    C. R. Banbury, V . J. Reddi, M. Lam, W. Fu, A. Fazel, J. Holleman, X. Huang, R. Hurtado, D. Kanter, A. Lokhmotov, D. Patterson, D. Pau, J. sun Seo, J. Sieracki, U. Thakker, M. Verhelst, and P. Yadav, “Bench- marking tinyml systems: Challenges and direction,” 2021

    2021

  12. [20]

    Tinyllava: A framework of small-scale large multimodal models,

    B. Zhou, Y . Hu, X. Weng, J. Jia, J. Luo, X. Liu, J. Wu, and L. Huang, “Tinyllava: A framework of small-scale large multimodal models,”arXiv preprint arXiv:2402.14289, 2024

    2024 arXiv

  13. [21]

    Rokid AI & AR glasses—redefining reality

    Rokid, “Rokid AI & AR glasses—redefining reality.” Product page,

  14. [22]

    Project Astra: A research prototype exploring the future of AI assistants

    Google DeepMind, “Project Astra: A research prototype exploring the future of AI assistants.” Google DeepMind, 2024

    2024

  15. [23]

    Google vs meta smart glasses: Which AI frames are better

    Geeky Gadgets, “Google vs meta smart glasses: Which AI frames are better.” Geeky Gadgets, 2026. Accessed 2026-05-11

    2026

  16. [24]

    Edge cloud offloading algorithms: Issues, methods, and perspectives,

    J. Wang, J. Pan, F. Esposito, P. Calyam, Z. Yang, and P. Mohapatra, “Edge cloud offloading algorithms: Issues, methods, and perspectives,” ACM Computing Surveys (CSUR), vol. 52, no. 1, pp. 1–23, 2019

    2019

  17. [25]

    Cosmos:computation offloading as a service for mobile devices,

    C. Shi, K. Habak, P. Pandurangan, M. Ammar, M. Naik, and E. Zegura, “Cosmos:computation offloading as a service for mobile devices,” in Proceedings of the 15th ACM International Symposium on Mobile ad Hoc Networking and Computing, pp. 287–296, 2014

    2014

  18. [26]

    To offload or not to offload? the bandwidth and energy costs of mobile cloud comput- ing,

    M. V . Barbera, S. Kosta, A. Mei, and J. Stefa, “To offload or not to offload? the bandwidth and energy costs of mobile cloud comput- ing,” inIEEE Conference on Computer Communications (INFOCOM), pp. 1285–1293, 2013

    2013

  19. [27]

    Images and vision,

    OpenAI, “Images and vision,” 2026. Accessed: 2026-04-01

    2026

  20. [28]

    Understand and count tokens — gemini api,

    Google, “Understand and count tokens — gemini api,” 2026. Accessed: 2026-04-01

    2026

  21. [29]

    A-vit: Adaptive tokens for efficient vision transformer,

    H. Yin, A. Vahdat, J. M. Alvarez, A. Mallya, J. Kautz, and P. Molchanov, “A-vit: Adaptive tokens for efficient vision transformer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pp. 10809–10818, 2022

    2022

  22. [30]

    Not all patches are what you need: Expediting vision transformers via token reorganiza- tions,

    Y . Liang, C. Ge, Z. Tong, Y . Song, J. Wang, and P. Xie, “Not all patches are what you need: Expediting vision transformers via token reorganiza- tions,” inInternational Conference on Learning Representations, 2022

    2022

  23. [31]

    Token merging: Your vit but faster,

    D. Bolya, C.-Y . Fu, X. Dai, P. Zhang, C. Feichtenhofer, and J. Hoffman, “Token merging: Your vit but faster,” inInternational Conference on Learning Representations, 2023

    2023

  24. [32]

    Elf: Accelerate high-resolution mobile deep vision with content-aware parallel offloading,

    W. Zhang, Z. He, L. Liu, Z. Jia, Y . Liu, M. Gruteser, D. Raychaudhuri, and Y . Zhang, “Elf: Accelerate high-resolution mobile deep vision with content-aware parallel offloading,” inProceedings of the 27th Annual International Conference on Mobile Computing and Networking, pp....

    2021

  25. [33]

    Accumo: Accuracy- centric multitask offloading in edge-assisted mobile augmented reality,

    Z. J. Kong, Q. Xu, J. Meng, and Y . C. Hu, “Accumo: Accuracy- centric multitask offloading in edge-assisted mobile augmented reality,” inProceedings of the 29th Annual International Conference on Mobile Computing and Networking, pp. 1–16, 2023

    2023

  26. [34]

    Deep contextualized compressive offloading for images,

    B. Chen, Z. Yan, H. Guo, Z. Yang, A. Ali-Eldin, P. Shenoy, and K. Nahrstedt, “Deep contextualized compressive offloading for images,” inProceedings of the 19th ACM Conference on Embedded Networked Sensor Systems, pp. 467–473, 2021

    2021

  27. [35]

    Towards wearable cognitive assistance,

    K. Ha, Z. Chen, W. Hu, W. Richter, P. Pillai, and M. Satyanarayanan, “Towards wearable cognitive assistance,” inProceedings of the 12th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys), pp. 68–81, 2014

    2014

  28. [36]

    Deep learning with edge computing: A review,

    J. Chen and X. Ran, “Deep learning with edge computing: A review,” Proceedings of the IEEE, vol. 107, no. 8, pp. 1655–1674, 2019

    2019

  29. [37]

    Color-to-grayscale: Does the method matter in image recognition?,

    C. Kanan and G. W. Cottrell, “Color-to-grayscale: Does the method matter in image recognition?,”PLOS ONE, vol. 7, no. 1, p. e29740, 2012

    2012

  30. [38]

    The jpeg still picture compression standard,

    G. K. Wallace, “The jpeg still picture compression standard,”Commu- nications of the ACM, vol. 34, no. 4, pp. 30–44, 1991

    1991

  31. [39]

    Learning to resize images for computer vision tasks,

    H. Talebi and P. Milanfar, “Learning to resize images for computer vision tasks,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 497–506, 10 2021

    2021

  32. [40]

    VQA-MHUG: A gaze dataset to study multimodal neural attention in visual question answering,

    E. Sood, F. K ¨ogel, F. Strohm, P. Dhar, and A. Bulling, “VQA-MHUG: A gaze dataset to study multimodal neural attention in visual question answering,” inProceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pp. 27–43, 2021

    2021

  33. [41]

    Eye gaze tells you where to compute: Gaze-driven efficient vlms,

    Q. Chen and J. Qi, “Eye gaze tells you where to compute: Gaze-driven efficient vlms,” 2025

    2025

  34. [42]

    Saliency detection: A spectral residual approach,

    X. Hou and L. Zhang, “Saliency detection: A spectral residual approach,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007

    2007

  35. [43]

    Saliency driven perceptual image compression,

    Y . Patel, S. Appalaraju, and R. Manmatha, “Saliency driven perceptual image compression,” inProceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pp. 227–236, 2021

    2021

  36. [44]

    Visual cropping improves zero-shot question answering of multimodal large language models,

    J. Zhang, M. Khayatkhoei, P. Chhikara, and F. Ilievski, “Visual cropping improves zero-shot question answering of multimodal large language models,” inR0-FoMo:Robustness of Few-shot and Zero-shot Learning in Large Foundation Models, 2023

    2023

  37. [45]

    Edge assisted real-time object detection for mobile augmented reality,

    L. Liu, H. Li, and M. Gruteser, “Edge assisted real-time object detection for mobile augmented reality,” inThe 25th Annual International Con- ference on Mobile Computing and Networking (MobiCom), pp. 1–16, 2019

    2019

  38. [46]

    Glimpse: Continuous, real-time object recognition on mobile devices,

    T. Y .-H. Chen, L. Ravindranath, S. Deng, P. Bahl, and H. Balakrishnan, “Glimpse: Continuous, real-time object recognition on mobile devices,” inProceedings of the 13th ACM Conference on Embedded Networked Sensor Systems (SenSys), pp. 155–168, 2015

    2015

  39. [47]

    Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering,

    Y . Goyal, T. Khot, D. Summers-Stay, D. Batra, and D. Parikh, “Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering,” inConference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  40. [48]

    MMBench: Is your multi- modal model an all-around player?,

    Y . Liu, H. Duan, Y . Zhang, B. Li, S. Zhang, W. Zhao, Y . Yuan, J. Wang, C. He, Z. Liu, K. Chen, and D. Lin, “MMBench: Is your multi- modal model an all-around player?,” inProceedings of the European Conference on Computer Vision (ECCV), 2024

    2024

  41. [49]

    MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI,

    X. Yue, Y . Ni, K. Zhang, T. Zheng, R. Liu, G. Zhang, S. Stevens, D. Jiang, W. Ren, Y . Sun, C. Wei, B. Yu, R. Yuan, R. Sun, M. Yin, B. Zheng, Z. Yang, Y . Liu, W. Huang, H. Sun, Y . Su, and W. Chen, “MMMU: A massive multi-discipline multimodal understanding and reasoning benc...

    2024

  42. [50]

    MathVista: Evaluating mathematical reasoning of foundation models in visual contexts,

    P. Lu, H. Bansal, T. Xia, J. Liu, C. Li, H. Hajishirzi, H. Cheng, K.-W. Chang, M. Galley, and J. Gao, “MathVista: Evaluating mathematical reasoning of foundation models in visual contexts,” inInternational Conference on Learning Representations (ICLR), 2024

    2024

  43. [51]

    HoloAssist: An egocentric human interaction dataset for interactive AI assistants in the real world,

    X. Wang, T. Kwon, M. Rad, B. Pan, I. Chakraborty, S. Andrist, D. Bohus, A. Feniello, B. Tekin, F. V . Frujeri, N. Joshi, and M. Pollefeys, “HoloAssist: An egocentric human interaction dataset for interactive AI assistants in the real world,” inProceedings of the IEEE/CVF Inter...

    2023

  44. [52]

    WearVQA: A visual question answering benchmark for wearables in egocentric authentic real-world scenarios,

    E. Chang, Z. Huang, Y . Liao, S. R. Bhavsar, A. Param, T. Stark,et al., “WearVQA: A visual question answering benchmark for wearables in egocentric authentic real-world scenarios,” inAdvances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  45. [53]

    SuperGlasses: Benchmarking vision language models as intelligent agents for AI smart glasses,

    Z. Jiang, X. Yuan, H. Qu, S. Lin, K. Liu, W. Fan, and Q. Li, “SuperGlasses: Benchmarking vision language models as intelligent agents for AI smart glasses,”arXiv preprint arXiv:2602.22683, 2026

    2026 arXiv

  46. [54]

    CRAG-MM: Multi-modal multi-turn comprehen- sive RAG benchmark,

    J. Wang, X. Yang, K. Sun, P. Suresh, S. Sharma, A. Czyzewski, D. Andersen,et al., “CRAG-MM: Multi-modal multi-turn comprehen- sive RAG benchmark,”arXiv preprint arXiv:2510.26160, 2025

    2025

  47. [55]

    OK-VQA: A visual question answering benchmark requiring external knowledge,

    K. Marino, M. Rastegari, A. Farhadi, and R. Mottaghi, “OK-VQA: A visual question answering benchmark requiring external knowledge,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3195–3204, 2019

    2019

  48. [56]

    MISAR: A multimodal instructional system with augmented reality,

    J. Bi, N. Nguyen, A. V osoughi, and C. Xu, “MISAR: A multimodal instructional system with augmented reality,” inProceedings of the 15 IEEE/CVF International Conference on Computer Vision (ICCV) Work- shop on AV4D, 2023

    2023

  49. [57]

    Guided Reality: Generating visually-enriched AR task guidance with LLMs and vision models,

    A. Y . Zhao, A. Gunturu, E. Y .-L. Do, and R. Suzuki, “Guided Reality: Generating visually-enriched AR task guidance with LLMs and vision models,” inProceedings of the 38th Annual ACM Symposium on User Interface Software and Technology (UIST), 2025

    2025

  50. [58]

    EmBARDiment: An embodied AI agent for productivity in XR,

    R. Bovo, S. Abreu, K. Ahuja, E. J. Gonzalez, L.-T. Cheng, and M. Gonzalez-Franco, “EmBARDiment: An embodied AI agent for productivity in XR,”arXiv preprint arXiv:2408.08158, 2024

    2024 arXiv

  51. [59]

    GazePointAR: A context-aware multimodal voice assistant for pronoun disambiguation in wearable augmented reality,

    J. Lee, J. Wang, E. Brown, L. Chu, S. S. Rodriguez, and J. E. Froehlich, “GazePointAR: A context-aware multimodal voice assistant for pronoun disambiguation in wearable augmented reality,” inProceedings of the 2024 CHI Conference on Human Factors in Computing Systems (CHI), pp...

    2024

  52. [60]

    Designing memory- augmented AR agents for spatiotemporal reasoning in personalized task assistance,

    D. Choi, T. Kwon, D. Yang, H. Kim, and J. Yeo, “Designing memory- augmented AR agents for spatiotemporal reasoning in personalized task assistance,”arXiv preprint arXiv:2508.08774, 2025

    2025 arXiv

  53. [61]

    Cross-format retrieval- augmented generation in XR with LLMs for context-aware maintenance assistance,

    A. Nagy, Y . Spyridis, and V . Argyriou, “Cross-format retrieval- augmented generation in XR with LLMs for context-aware maintenance assistance,”arXiv preprint arXiv:2502.15604, 2025

    2025 arXiv

  54. [62]

    Exploring the use of VLMs for navigation assistance for people with blindness and low vision,

    Y . Li, Y . Zheng, G. Hamilton-Fletcher, M. Mezzavilla, Y . Wang, S. Ran- gan, M. Porfiri, Z. Yu, and J.-R. Rizzo, “Exploring the use of VLMs for navigation assistance for people with blindness and low vision,”arXiv preprint arXiv:2603.15624, 2026

    2026

  55. [63]

    BLaVe-CoT: Consistency-aware visual question answering for blind and low vision users,

    W. Cheng and Z. Ruan, “BLaVe-CoT: Consistency-aware visual question answering for blind and low vision users,”arXiv preprint arXiv:2509.06010, 2025

    2025 arXiv

  56. [64]

    Objective assessment of the webp image coding algorithm,

    G. Ginesu, R. Pintus, and D. D. Giusto, “Objective assessment of the webp image coding algorithm,”Signal Processing: Image Communica- tion, vol. 27, no. 8, pp. 867–877, 2012

    2012

  57. [65]

    Vision - claude api docs,

    Anthropic, “Vision - claude api docs,” 2026. Accessed: 2026-04-01

    2026

  58. [66]

    Image understanding — gemini api,

    Google, “Image understanding — gemini api,” 2026. Accessed: 2026- 04-01

    2026

  59. [67]

    Image interpolation and resam- pling,

    P. Th ´evenaz, T. Blu, and M. Unser, “Image interpolation and resam- pling,” inHandbook of Medical Imaging, Processing and Analysis(I. N. Bankman, ed.), pp. 393–420, Academic Press, 2000

    2000

  60. [68]

    Colorbench: Can vlms see and understand the colorful world? a comprehensive benchmark for color perception, reasoning, and robustness,

    Y . Liang, M. Li, C. Fan, Z. Li, D. Nguyen, K. Cobbina, S. Bhardwaj, J. Chen, F. Liu, and T. Zhou, “Colorbench: Can vlms see and understand the colorful world? a comprehensive benchmark for color perception, reasoning, and robustness,”Advances in Neural Information Processing ...

    2026

  61. [69]

    V oila-a: Aligning vision-language models with user's gaze attention,

    K. Yan, Z. Wang, L. Ji, Y . Wang, N. Duan, and S. Ma, “V oila-a: Aligning vision-language models with user's gaze attention,” inAdvances in Neural Information Processing Systems(A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, eds.), vol. 37, ...

    1918

  62. [70]

    EgoVQA—an egocentric video question answering benchmark dataset,

    C. Fan, “EgoVQA—an egocentric video question answering benchmark dataset,” inProceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pp. 4359–4366, 2019

    2019

  63. [71]

    Neural machine translation of rare words with subword units,

    R. Sennrich, B. Haddow, and A. Birch, “Neural machine translation of rare words with subword units,” inProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), (Berlin, Germany), pp. 1715–1725, Association for Computation...

    2016

  64. [72]

    What are tokens and how to count them?

    OpenAI, “What are tokens and how to count them?.” OpenAI Help Center, 2026. Accessed 2026-05-14

    2026

  65. [73]

    Saliency based image crop- ping,

    E. Ardizzone, A. Bruno, and G. Mazzola, “Saliency based image crop- ping,” inInternational Conference on Image Analysis and Processing, pp. 773–782, Springer, 2013

    2013

  66. [74]

    Benchmarking deep learning models for object detection on edge computing devices,

    D. K. Alqahtani, M. A. Cheema, and A. N. Toosi, “Benchmarking deep learning models for object detection on edge computing devices,” inService-Oriented Computing(W. Gaaloul, M. Sheng, Q. Yu, and S. Yangui, eds.), (Singapore), pp. 142–150, Springer Nature Singapore, 2025

    2025

  67. [75]

    Region-of-interest extraction method to increase object-detection performance in remote monitoring system,

    H.-G. Jeon and K.-H. Lee, “Region-of-interest extraction method to increase object-detection performance in remote monitoring system,” Applied Sciences, vol. 15, no. 10, p. 5328, 2025

    2025

  68. [76]

    Improving automatic VQA evaluation using large language models,

    O. Ma ˜nas, B. Krojer, and A. Agrawal, “Improving automatic VQA evaluation using large language models,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 4171–4179, 2024

    2024

  69. [77]

    Mind the uncertainty in human disagreement: Evaluating discrepancies between model predictions and human responses in vqa,

    J. Lan, D. Frassinelli, and B. Plank, “Mind the uncertainty in human disagreement: Evaluating discrepancies between model predictions and human responses in vqa,” 2024

    2024

  70. [78]

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica, “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 36, 2023

    2023

  71. [79]

    G-Eval: NLG evaluation using GPT-4 with better human alignment,

    Y . Liu, D. Iter, Y . Xu, S. Wang, R. Xu, and C. Zhu, “G-Eval: NLG evaluation using GPT-4 with better human alignment,” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 2511–2522, 2023

    2023

  72. [80]

    Chatar: Conversation support using large language model and augmented reality,

    Y . Fujimoto, “Chatar: Conversation support using large language model and augmented reality,” 2025

    2025

  73. [81]

    Next-generation networking and edge computing for mixed reality real-time interactive systems,

    S. Shannigrahi, S. Mastorakis, and F. R. Ortega, “Next-generation networking and edge computing for mixed reality real-time interactive systems,” in2020 IEEE International Conference on Communications Workshops (ICC Workshops), pp. 1–6, 2020

    2020

  74. [82]

    An egocentric vision-language model based portable real-time smart assistant,

    Y . Huang, J. Xu, B. Pei, Y . He, G. Chen, M. Zhang, L. Yang, Z. Nie, J. Liu, G. Fan, D. Lin, F. Fang, K. Li, C. Yuan, Y . Wang, X. Chen, Y . Wang, Y . Qiao, and L. Wang, “An egocentric vision-language model based portable real-time smart assistant,”Proceedings of the ACM on I...

    2025

  75. [83]

    Teaching LLMs to see and guide: Context-aware real-time assistance in augmented reality,

    M. Qorbani, K. Paynabar, and M. Moghaddam, “Teaching LLMs to see and guide: Context-aware real-time assistance in augmented reality,” arXiv preprint arXiv:2511.00730, 2025

    2025

  76. [84]

    Gartner predicts 80% of enterprise software and applications will be multimodal by 2030, up from less than 10% in 2024,

    Gartner, “Gartner predicts 80% of enterprise software and applications will be multimodal by 2030, up from less than 10% in 2024,” July 2025

    2024

  77. [85]

    Deep learning-based object detection in augmented reality: A systematic review,

    Y . Ghasemi, H. Jeong, S. H. Choi, K.-B. Park, and J. Y . Lee, “Deep learning-based object detection in augmented reality: A systematic review,”Computers in Industry, vol. 139, p. 103661, 2022

    2022

  78. [86]

    Realtime API with WebRTC,

    OpenAI, “Realtime API with WebRTC,” 2026. Accessed: 2026

    2026

  79. [87]

    Realtime API with WebSocket,

    OpenAI, “Realtime API with WebSocket,” 2026. Accessed: 2026

    2026

  80. [88]

    Bench- marking the effects of operating system interference on extreme-scale parallel machines,

    P. Beckman, K. Iskra, K. Yoshii, S. Coghlan, and A. Nataraj, “Bench- marking the effects of operating system interference on extreme-scale parallel machines,”Cluster Computing, vol. 11, no. 1, pp. 3–16, 2008

    2008

  81. [89]

    Understanding the causes of performance variability in HPC workloads,

    D. Skinner and W. Kramer, “Understanding the causes of performance variability in HPC workloads,” inProceedings of the IEEE International Workload Characterization Symposium, pp. 137–149, IEEE, 2005

    2005

  82. [90]

    Overhead mea- surement noise in different runtime environments,

    D. G. Reichelt, R. Jung, and A. van Hoorn, “Overhead mea- surement noise in different runtime environments,”arXiv preprint arXiv:2411.05491, 2024

    2024 arXiv

  83. [91]

    Using microbenchmarks to evaluate system performance,

    B. N. Bershad, R. P. Draves, and A. Forin, “Using microbenchmarks to evaluate system performance,” inProceedings of the Third Workshop on Workstation Operating Systems (WWOS-III), pp. 148–153, 1992

    1992

  84. [92]

    Beyond inference: Performance analysis of DNN server overheads for computer vision,

    A. F. AbouElhamayed, S. Balle, D. Singh, and M. S. Abdelfattah, “Beyond inference: Performance analysis of DNN server overheads for computer vision,” inProceedings of the 61st ACM/IEEE Design Automation Conference (DAC), 2024

    2024

  85. [93]

    EdgeYOLO: An edge-real- time object detector,

    S. Liu, J. Zha, J. Sun, Z. Li, and G. Wang, “EdgeYOLO: An edge-real- time object detector,” in2023 42nd Chinese Control Conference (CCC), pp. 7507–7512, IEEE, 2023

    2023

  86. [94]

    Images and vision — calculating image tokens

    OpenAI, “Images and vision — calculating image tokens.” OpenAI API Documentation, 2026. Accessed: 2026-05-28

    2026

  87. [95]

    Per-part media resolution (Gemini 3 only)

    Google, “Per-part media resolution (Gemini 3 only).” https://ai.google. dev/gemini-api/docs/media-resolution, 2026. Accessed: 2026-05-28

    2026

  88. [96]

    Gemini 3 Flash — model documentation

    Google Cloud, “Gemini 3 Flash — model documentation.” https://docs. cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-flash,

  89. [97]

    Gemini 3 developer guide — generateContent API

    Google, “Gemini 3 developer guide — generateContent API.” https: //ai.google.dev/gemini-api/docs/gemini-3, 2026. Accessed: 2026-05-28. APPENDIX A. Hardware and Network Setup The experiments were performed on a laptop equipped with an Intel Core i7-8750H CPU at 2.2 GHz, 16 GB o...

    2026

  90. [2026]

    Accessed: 2026-05-28

    Default resolution tokens: 1120. Accessed: 2026-05-28

    2026

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.