TALSC: Timeliness-Aware Large-Small VLM Collaboration for Infrastructure-Assisted Autonomous Driving

Bo Ai; Mengmeng Zhu; Wei Chen; Yuxuan Sun

REVIEW 4 major objections 5 minor 30 references

TALSC: Timeliness-Aware Large-Small VLM Collaboration for Infrastructure-Assisted Autonomous Driving

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper argues that a timeliness-aware online scheduler, choosing per task whether to invoke a small onboard or large edge vision-language model, how many visual tokens to send, and how much bandwidth to allocate, can maximize long-run t

desk verdict Plausible Lyapunov scheduler for VLM offloading, but the 12.6% gain is an artifact of evaluating on the same fitted surrogate it maximizes. read the letter →

arxiv 2608.01998 v1 pith:FDU55U6P submitted 2026-08-03 cs.DC cs.AIcs.NI

Mengmeng Zhu , Yuxuan Sun , Wei Chen , Bo Ai This is my paper

classification cs.DCcs.AIcs.NI

keywords AgeofInformationVision-LanguageModelsAutonomousDrivingEdgeComputingLyapunovoptimizationOnlineschedulingTimelinessmetricTokenlengthadaptation

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

A self-driving vehicle running a small vision-language model on board can offload hard perception tasks to a large model on an edge server, but that takes time and bandwidth, and the data goes stale quickly. This paper proposes TALSC, an online scheduling algorithm that for each task decides local-versus-edge inference, the number of visual tokens to transmit, and the bandwidth share, with the goal of maximizing a long-term timeliness metric that combines information age and token length. The algorithm handles two practical difficulties: decisions made now affect the freshness of future tasks, and the number of output tokens is unknown at scheduling time. The paper claims a guaranteed gap to the offline optimum and reports up to a 12.6% normalized improvement in Micro-F1 over the best baseline in simulation. If correct, this gives a principled way to trade accuracy against freshness in infrastructure-assisted autonomous driving.

What carries the argument

The load-bearing constructs are (i) the timeliness metric Phi_x(h,L) = alpha_x - beta_x e^{-mu_x L} + delta_x e^{-lambda_x h}, a double-exponential fit that maps AoI h and visual token length L to expected task performance measured by Micro-F1; (ii) the Lyapunov drift-plus-estimated-penalty scheduler, which at each task minimizes x_m Q_m b_m - V * Utilde_m over a discrete enumeration of collaboration decisions and token lengths, where Q_m is a virtual queue tracking bandwidth overuse; and (iii) a Taylor-expansion estimation step that predicts the next AoI using the expected output token count. Together they convert a long-term stochastic optimization with delayed effects into a per-task myop

What would settle it

A concrete test: take a held-out set of driving images and at least two different VLM sizes, fit the double-exponential Phi from Eq. (18) on one split, and check whether it predicts Micro-F1 on the other split within, say, 0.05 absolute error. If the fit fails, or if applying the same scheduler to a different task (e.g., motion prediction) shows no gain, the 12.6% improvement is an artifact of the surrogate. A separate check: recompute Eq. (8) with the proper normalization of the AoI average; if the estimate changes materially, the drift-plus-penalty guarantee in Theorem 1 may not apply to the

Watch

Extended reading notes

Core claim

The central claim is that VLM task performance in infrastructure-assisted driving can be captured by a parametric double-exponential surface Phi(h,L) that couples the age of information h and visual token length L, and that an online scheduler maximizing the long-run average of this surface — while bounding bandwidth via a virtual queue and latency via a hard deadline — achieves a guaranteed fraction of the offline optimum. Specifically, Theorem 1 shows that the accumulated timeliness of TALSC is at least the offline optimum minus O(M/V) plus a bounded error term, while the average bandwidth violation is O(sqrt(V)). The empirical case study fits Phi using Qwen2.5-VL-3B (local) and Qwen2.5-VL

Load-bearing premise

Everything rests on the fitted double-exponential timeliness metric Phi(h,L) being an accurate, transferable model of real VLM task performance; it is fit on 400 nuScenes images with two Qwen models and is never validated on held-out data or with error bars.

Editorial extensions

If this is right

For each arriving task, TALSC makes a one-pass decision on local-versus-edge inference, visual token length, and bandwidth without needing future task information beyond statistical estimates.
The long-term bandwidth budget is respected in the sense of a virtual-queue bound (Eq. 14), so the policy avoids exceeding the average bandwidth constraint.
Hard latency deadlines are enforced directly: any (x, L) pair whose estimated total latency exceeds T_req is pruned during enumeration, ensuring the safety-critical deadline per task.
The fitted empirical surface shows diminishing returns in both token length and AoI, implying the optimal operating point is interior: neither always-edge nor always-local dominates across channel and latency conditions.
If the surrogate Phi is accurate, the reported 12.6% normalized Micro-F1 gain indicates a practically meaningful improvement in poor channel conditions, where adaptive token length keeps data useful without overflowing bandwidth.

Reading between the lines

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

The drift-plus-estimated-penalty template could generalize to other semantic compression knobs, such as quantization bit-width or region-of-interest cropping, provided a timeliness surrogate exists for those knobs.
The double-exponential form of Phi is likely task- and model-specific; adapting TALSC to a new deployment would require recalibration or a meta-model that predicts the fitting parameters from model characteristics and dataset statistics.
The formulation treats sensory data as given each period; a natural extension is to couple scheduling with sensing decisions, e.g., whether to skip a frame or choose a sensor subset, which would change the AoI reset dynamics entirely.
Equation (8)'s estimate of the mean AoI is written as an unnormalized integral with units of time squared, so the practical tuning of the algorithm and the constant C in the guarantee could shift if that estimator is replaced by a properly normalized version.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Couple things you should know. This is a credible Lyapunov scheduling paper for a real problem—deciding when a vehicle should use its small VLM or offload to an edge large VLM, and how many visual tokens to send—but the headline 12.6% Micro-F1 gain is computed with the same fitted surrogate function that the scheduler maximizes, not with real VLM outputs. The improvement is therefore circular. There's also a genuine units bug in Eq (8): the estimated average AoI is written as an interval integral, giving units of time^2, which can't be input to Phi.

What's new and worth keeping: the problem formulation coupling AoI, token length, and model choice is a legitimate extension of the edge-assisted perception literature. The AoI evolution model for VLM inference is sensible, and the latency model for token-based transmission and inference is reasonable. The Lyapunov drift-plus-estimated-penalty algorithm is standard but cleanly applied, and the offline regret bound is a nice add-on, even if the proof is compressed.

The soft spots are real but fixable. First, the circular evaluation: all the Micro-F1 numbers in Fig. 4 come from evaluating the fitted double-exponential Phi (Eq. 18) on each scheduled task. Since TALSC maximizes exactly that Phi, the comparison just shows the algorithm is better at optimizing its own objective. The fit itself is based on 400 nuScenes images from 10 clips, with no held-out validation, error bars, or test of the double-exponential assumption. So the 12.6% claim tells us nothing about real VLM accuracy. Second, the unit error in Eq. (8) is load-bearing—it feeds into the estimated penalty in P3 and likely affects the scheduling choices. Third, the Theorem 1 proof is sketchy: the bounding constants C and z0 aren't explicitly defined with all their dependencies, and the step from the drift-plus-penalty bound to the queue bound is compressed. The paper doesn't provide code or data, which makes verification harder, but the model and derivations are mostly transparent.

Who benefits: readers working on edge-assisted autonomous driving, AoI-based resource allocation, or efficient VLM deployment. It's not a bad paper; it's an unfinished one. The core idea and the optimization machinery are sound enough to warrant referee time. I'd suggest the editor send it to review with the expectation of major revision: rerun the experiments with actual VLM inference on held-out data, fix the units in Eq. (8), tighten the theorem statements, and either provide code or report error bars.

My bottom line: engage with it, but don't trust the empirical claims until they're redone.

Referee Report

4 major / 5 minor

Summary. TALSC proposes an online scheduler for infrastructure-assisted autonomous driving in which a vehicle's small VLM (SVLM) can collaborate with a large VLM (LVLM) at an edge server. The scheduler jointly decides the collaboration decision x_m, visual-token length L_m, and bandwidth b_m so as to maximize the long-term average of a timeliness metric that is modeled as a double-exponential function of AoI and token length, subject to a long-term bandwidth budget and a per-task hard deadline. The paper models VLM inference latency, derives a Lyapunov drift-plus-estimated-penalty algorithm, states a performance guarantee relative to an offline optimum, and reports simulation results, including an up-to-12.6% normalized Micro-F1 improvement over baselines, using a timeliness function fitted to 400 nuScenes images with Qwen2.5-VL-3B/72B.

Significance. If the central claims were established, the paper would be a useful step toward practical large-small VLM collaboration: it explicitly models the AoI/token-length/accuracy coupling, formulates a clean online scheduling problem, and applies a Lyapunov-style drift-plus-penalty method with a stated performance bound. The empirical latency modeling and the use of real Qwen-VLM workloads are also constructive. However, the paper's main claims are not currently supported. The average-AoI estimate in Eq. (8) is dimensionally wrong; the subproblem decomposition in Eq. (12) ignores the dependence of the objective on b through the AoI; the proof of Theorem 1 contains an invalid expectation step; and the simulation's 'Micro-F1' is the same fitted surrogate function that the scheduler maximizes, making the evaluation circular. These are load-bearing issues, not presentation concerns.

major comments (4)

[§III-A, Eq. (8)] The right-hand side of Eq. (8) is the unnormalized integral of AoI over the interval [t'_m, t'_{m+1}), not the interval average \bar h_m defined immediately above it. Its units are time^2, so substituting it into Φ(h,L) as an AoI value is dimensionally inconsistent. The correct average AoI is T_total,m(\tilde K_m)+|T_m|/2. Because \tilde h_m is used to form \tilde U_m and hence the P3 objective and Algorithm 1, the algorithm optimizes the wrong estimated quantity.
[§III-A, Eq. (12)] Eq. (12) defines b^(k)_m = argmin_b [Q_m b - VΔT Φ_1(\tilde h_m, ℓ_k)], but \tilde h_m depends on b through T_total,m via T_comm,m = D(L)/R(b_m). The objective is therefore nonlinear in b through both the explicit bandwidth cost and the AoI argument inside Φ. The displayed argmin is not a well-posed subproblem unless \tilde h_m is treated as fixed, and the text does not state such an approximation. Thus Algorithm 1's Step 5 is not implementable as written, and the claimed 'without loss of optimality' decomposition is invalid.
[§III-B, proof of Theorem 1, Eq. (15)] The proof asserts that E[y*_m] ≤ 0 implies E[Q_m y*_m] ≤ 0. This is false: Q_m and y*_m are generally correlated, and a policy satisfying a long-term average constraint may use positive y*_m in exactly the slots where Q_m is large. Standard Lyapunov arguments would compare with a stationary randomized policy or use a bound involving Q_m y*_m directly. As written, the proof of the performance guarantee in Eq. (13) and the queue bound in Eq. (14) is incomplete and the theorem is not established.
[§IV-C, Fig. 4 and Eq. (18)] The 'Micro-F1' reported in Fig. 4 is computed from the fitted timeliness function Φ(h,L) of Eq. (18) with the parameters of Table I. This is exactly the objective that P1/P3 maximize. The simulation therefore measures how well each policy optimizes the authors' own fitted surrogate, not the actual Micro-F1 of Qwen2.5-VL on scheduled tasks. No held-out validation, error bars, fit quality metrics, or comparison against the 400 empirical points is provided. Consequently the claimed up-to-12.6% normalized improvement is not evidence about real VLM task performance; it is a self-consistency check.

minor comments (5)

[Abstract] Typo: 'provides a guaranteed performance' should be 'provides a guaranteed performance bound' or 'provides performance guarantees.'
[Algorithm 1, line 15] Typo: 'algorith' should be 'algorithm.'
[§IV-A, Fig. 3 and Table I] The double-exponential model in Eq. (18) is asserted without reporting goodness-of-fit (e.g., R²), confidence intervals for the fitted parameters, or validation on held-out images. These are needed to assess whether the fitted surface is a reliable empirical model.
[§IV-C, Fig. 4] The 'normalized improvement' over the 'feasible performance range' is not defined. Please state the normalization formula explicitly.
[Reference [28]] The edgeVLM reference is incomplete: no venue, year, or arXiv identifier is given, despite the paper being compared with it as a baseline.

Circularity Check

1 steps flagged · score 6.0 of 10

Simulation's Micro-F1 is the same fitted surrogate the scheduler optimizes, so the 12.6% gain is circular and does not validate real VLM performance.

fitted input called prediction [Abstract; Section IV-A Eq. (18) and Table I; Section IV-C Fig. 4]
"we first conduct a case study to derive a fitted timeliness metric based on nuScenes dataset, and further show that TALSC outperforms baselines under various communication and computing settings, achieving up to a 12.6% normalized improvement in Micro-F1 score. ... The empirical timeliness data reveal the relationship among the task performance (measured by Micro-F1), AoI, and visual token length ... adopt a double exponential decay model to fit the data. Thus, for each task the timeliness metric is formulated as: Φ(h, L) = α_x − β_x · e^{−μ_x L} + δ_x e^{−λ_x h}. The fitted parameters are sum"

The scheduling objective P1 (Eq. 5a) maximizes the long-term average of Φ_{x_m}(h_t, L_m), and the online algorithm P3 (Eq. 11a) minimizes the drift-plus-penalty term containing \tilde{U}_m, which is explicitly built from Φ. The reported time-averaged Micro-F1 in Fig. 4 is never computed by running Qwen VLMs on the scheduled decisions; the only Micro-F1 model described in the paper is the fitted double-exponential Φ of Eq. 18 with the Table I parameters. Thus the evaluation metric is identical, by construction, to the reward that TALSC is designed to maximize. A policy that better fits this surrogate will trivially appear better on the same fitted curve, so the claimed 12.6% normalized improvement is an artifact of using the fitted function as both objective and outcome measure, not eviden

full rationale

The central circularity is that the paper fits a surrogate Φ(h,L) to micro-F1 data (Eq. 18, Table I), uses that same Φ as the reward in P1/P3, and then reports 'Micro-F1' performance in Fig. 4 without any independent evaluation on actual VLM outputs. This is the fitted-reward-as-evaluation pattern: the optimization and the metric coincide by construction, so the headline performance gain does not validate the algorithm against real VLM behavior. The Lyapunov drift-plus-penalty derivation (Theorem 1, Eq. 13) is mathematically independent of the surrogate and is not circular by itself; it provides a bound relative to the offline optimum of the same Φ-based objective. The empirical latency models (Table II) are also independent measurements. However, the paper's main simulation claim—that TALSC 'significantly outperforms existing baselines' in Micro-F1—reduces to the fitted Φ, because no alternative Micro-F1 computation is described. There is also a units inconsistency in Eq. (8) (the average AoI estimate is written as an unnormalized interval integral), but that is a correctness issue, not circularity. No load-bearing self-citation was found; references to Y. Sun's prior work are contextual and not used to justify the algorithm's optimality. Overall score 6: the central empirical claim partially reduces to a fit, though the online scheduling framework itself has independent content.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central claim depends on a heavily fitted empirical pipeline as much as on the algorithmic derivation. Ten parameters define the timeliness metric, five more define the latency models, and the Lyapunov weight V is unspecified. The double-exponential form of Phi is an ad hoc modeling choice, and the boundedness assumptions on estimation error are stated without verification. No fundamentally new physical or mathematical entities are introduced.

free parameters (5)

alpha_3B, beta_3B, mu_3B, delta_3B, lambda_3B (SVLM fit) = 0.3227, 0.1088, 0.0183, 0.3149, 0.0700
Fitted to nuScenes empirical Micro-F1 data; these coefficients shape the entire objective in Eq (18) and the evaluation in Fig 4.
alpha_72B, beta_72B, mu_72B, delta_72B, lambda_72B (LVLM fit) = 0.4600, 0.1106, 0.0081, 0.2429, 0.1369
Fitted to nuScenes empirical Micro-F1 data for the edge 72B model; used as the timeliness metric for offloaded tasks.
theta_per, tau_per = 0.0441 ms/token, 38.10 ms
Linear perception latency model fitted to Qwen2.5-VL-3B on an A4000 GPU; used in the total latency constraint.
theta_comp, tau_comp, tau_out (local and edge) = Table II values
Empirical computation latency model for local 3B and edge 72B inference, treated as exact in scheduling.
V (Lyapunov trade-off parameter) = not reported
Hand-chosen or tuned; controls the trade-off between objective maximization and bandwidth-queue stability, and appears in the performance bound.

assumptions (7)

domain assumption AoI resets on task completion and grows linearly with time
Section II-B, Eq (4); standard AoI model adopted from [12].
domain assumption Tasks arrive periodically with interval Delta T
Section II-A; VQA queries triggered by high-level driving decisions.
ad hoc to paper Timeliness Phi(h,L) has the double-exponential form of Eq (18)
Chosen to fit observed diminishing returns; not derived from any first-principles model.
domain assumption Expected output token count is available for estimation
Section III-A uses the statistical expectation of K_m to estimate latency and AoI, but no model, estimator, or empirical distribution is supplied.
domain assumption Estimation errors delta_m, epsilon_m, zeta_m are bounded
Section III-A, Eq (9) and Theorem 1 require finite bounds on estimation and Taylor remainder terms.
domain assumption Latency models are linear in L and K
Section IV-B fits T_per = theta L + tau and T_comp = theta L + tau + tau_out K and treats them as exact within scheduling.
standard math Phi is smooth enough for Taylor expansion with bounded remainder
Eq (7) uses a second-order Taylor expansion and assumes the remainder term epsilon_m is bounded.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TALSC: Timeliness-Aware Large-Small VLM Collaboration for Infrastructure-Assisted Autonomous Driving." pith.science (2026). https://pith.science/paper/FDU55U6P

@misc{pith2026260801998,
  author       = {Pith},
  title        = {Pith review of: TALSC: Timeliness-Aware Large-Small VLM Collaboration for Infrastructure-Assisted Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FDU55U6P}},
  note         = {Machine review of arXiv:2608.01998}
}

read the original abstract

The deployment of Vision-Language Models (VLMs) in autonomous driving (AD) systems is constrained by on-board computing power, restricting vehicles to small VLMs (SVLMs) with limited perception and reasoning capabilities. Infrastructure-assisted AD alleviates this resource constraint by enabling collaboration with large VLMs (LVLMs) at edge servers. However, in dynamic vehicular environments, the utility of sensory data for downstream tasks decays rapidly, making timeliness of information a critical concern. To balance the accuracy gains of LVLMs with their latency-induced timeliness degradation, we develop a Timeliness-Aware Large-Small VLM Collaboration (TALSC) framework. Specifically, we first model the Age of Information (AoI) evolution for VLM inference and characterize the coupling among AoI, token length, and task performance to formulate a general timeliness metric. Building on this, we propose the TALSC online scheduling algorithm. Since scheduling decisions have a delayed impact on future timeliness metric and the output token number is unknown at scheduling time, we design a Lyapunov drift-plus-estimated-penalty algorithm and provides a guaranteed performance. In simulation, we first conduct a case study to derive a fitted timeliness metric based on nuScenes dataset, and further show that TALSC outperforms baselines under various communication and computing settings, achieving up to a 12.6\% normalized improvement in Micro-F1 score compared with the best-performing baseline.

Figures

Figures reproduced from arXiv: 2608.01998 by the authors.

**Figure 2.** Evolution of AoI and the proposed timeliness metric. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 4.** Output token number dynamics and performance comparison under [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 2 linked inside Pith

[1]

Open- DriveVLA: Towards end-to-end autonomous driving with large vision language action model,

X. Zhou, X. Han, F. Yang, Y . Ma, V . Tresp, and A. Knoll, “Open- DriveVLA: Towards end-to-end autonomous driving with large vision language action model,” inProc. AAAI Conf. Artif. Intell. (AAAI), Mar 2026, pp. 13 782–13 790

2026
[2]

VLM-MPC: Model predictive controller augmented vision language model for autonomous driving,

K. Long, H. Shi, J. Liu, C. Xiao, and X. Li, “VLM-MPC: Model predictive controller augmented vision language model for autonomous driving,”Transportation Research Part C: Emerging Technologies, vol. 183, p. 105487, 2026

2026
[3]

A cloud–edge collabora- tive architecture for multimodal LLM-based advanced driver assistance systems in iot networks,

Y . Hu, D. Ye, J. Kang, M. Wu, and R. Yu, “A cloud–edge collabora- tive architecture for multimodal LLM-based advanced driver assistance systems in iot networks,”IEEE Internet Things J., vol. 12, no. 10, pp. 13 208–13 221, 2025

2025
[4]

guided mllm reasoning: Enhancing mllm with knowledge and visual notes for visual question answering,

W. Fang, Q. Wu, J. Chen, and Y . Xue, “guided mllm reasoning: Enhancing mllm with knowledge and visual notes for visual question answering,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 19 597–19 607

2025
[5]

EM-VLM4AD: Multi- frame, lightweight & efficient vision-language models for question answering in autonomous driving,

A. Gopalkrishnan, R. Greer, and M. Trivedi, “EM-VLM4AD: Multi- frame, lightweight & efficient vision-language models for question answering in autonomous driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) Workshops. IEEE, 2024, pp. 4552–4561

2024
[6]

A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms,

W. Zhao, Y . Han, J. Tang, Z. Li, Y . Song, K. Wang, Z. Wang, and Y . You, “A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 19 814–19 824

2025
[7]

Making large vision language models to be good few-shot learners,

F. Liu, W. Cai, J. Huo, C. Zhang, D. Chen, and J. Zhou, “Making large vision language models to be good few-shot learners,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2025, pp. 5415–5423

2025
[8]

Vision-language models for edge networks: A comprehensive survey,

A. Sharshar, L. U. Khan, W. Ullah, and M. Guizani, “Vision-language models for edge networks: A comprehensive survey,”IEEE Internet Things J., vol. 12, no. 16, pp. 32 701–32 724, 2025

2025

Show all 30 references

[9]

Task-oriented wireless communications for collaborative perception in intelligent un- manned systems,

S. Zhou, Y . Jia, R. Mao, Z. Nan, Y . Sun, and Z. Niu, “Task-oriented wireless communications for collaborative perception in intelligent un- manned systems,”IEEE Network, vol. 38, no. 6, pp. 21–28, 2024

2024
[10]

Timeliness-oriented asynchronous task offloading in UA V-edge- computing systems,

X. Qin, Y . Li, N. Ma, Y . Zhang, K. Han, L. Meng, and P. Zhang, “Timeliness-oriented asynchronous task offloading in UA V-edge- computing systems,”IEEE Trans. Netw. Sci. Eng., vol. 11, no. 1, pp. 900–912, 2023

2023
[11]

Optimizing task-specific timeliness with edge-assisted scheduling for status update,

J. Sun, L. Wang, Z. Nan, Y . Sun, S. Zhou, and Z. Niu, “Optimizing task-specific timeliness with edge-assisted scheduling for status update,” IEEE Journal on Selected Areas in Information Theory, vol. 4, pp. 624– 638, 2023

2023
[12]

Real-time status: How often should one update?

S. Kaul, R. Yates, and M. Gruteser, “Real-time status: How often should one update?” inProc. IEEE Conf. Comput. Commun. (INFOCOM). IEEE, 2012, pp. 2731–2735

2012
[13]

Timely monitoring of dynamic sources with observations from multiple wireless sensors,

A. E. Kalør and P. Popovski, “Timely monitoring of dynamic sources with observations from multiple wireless sensors,”IEEE/ACM Trans. Netw., vol. 31, no. 3, pp. 1263–1276, 2022

2022
[14]

Minimizing age of usage information for capturing freshness and usability of correlated data in edge computing enabled IoT systems,

X. Xie and H. Wang, “Minimizing age of usage information for capturing freshness and usability of correlated data in edge computing enabled IoT systems,”IEEE Trans. Mobile Comput., vol. 23, no. 5, pp. 5644–5659, 2023

2023
[15]

Minimizing task-oriented age of information for remote monitoring with pre-identification,

S. Gan, X. Wang, C. Xu, and X. Chen, “Minimizing task-oriented age of information for remote monitoring with pre-identification,”arXiv preprint arXiv:2508.14575, 2025

2025 arXiv
[16]

Granularity-aware hyperbolic representation for text-based person search,

C. Qi, X. Yang, N. Wang, and X. Gao, “Granularity-aware hyperbolic representation for text-based person search,”IEEE Trans. Inf. Forensics Secur., vol. 20, pp. 5745–5757, 2025

2025
[17]

Thinking in granularity: Dynamic quantization for image super-resolution by intriguing multi-granularity clues,

M. Wang, Z. Zhang, F. Li, K. Xu, K. Miao, and M. Wang, “Thinking in granularity: Dynamic quantization for image super-resolution by intriguing multi-granularity clues,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2025, pp. 7826–7834

2025
[18]

Granularity-adaptive spatial evidence tokenization for video question answering,

H. Jiang, Y . Jin, Z. Sun, K. Xu, K. Xu, L. Chen, Y . Song, K. Gai, and Y . Mu, “Granularity-adaptive spatial evidence tokenization for video question answering,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2025, pp. 3976–3984

2025
[19]

Tokenselect: Efficient long-context inference and length extrapolation for llms via dynamic token-level kv cache selection,

W. Wu, Z. Pan, K. Fu, C. Wang, L. Chen, Y . Bai, T. Wang, Z. Wang, and H. Xiong, “Tokenselect: Efficient long-context inference and length extrapolation for llms via dynamic token-level kv cache selection,” in Proceedings of the 2025 Conference on Empirical Methods in Natural ...

2025
[20]

Sliminfer: Accelerating long-context llm inference via dynamic token pruning,

L. Long, R. Yang, Y . Huang, D. Hui, A. Zhou, and J. Yang, “Sliminfer: Accelerating long-context llm inference via dynamic token pruning,” in Proc. AAAI Conf. Artif. Intell. (AAAI), 2026, pp. 32 284–32 292

2026
[21]

V2X-VLM: End-to-end V2V cooperative autonomous driving through large vision-language models,

J. You, Z. Jiang, Z. Huang, H. Shi, R. Gan, K. Wu, X. Cheng, X. Li, and B. Ran, “V2X-VLM: End-to-end V2V cooperative autonomous driving through large vision-language models,”Transportation Research Part C: Emerging Technologies, vol. 183, p. 105457, 2026

2026
[22]

Laecips: large vision model assisted adaptive edge-cloud collaboration for iot-based perception system,

S. Hu, R. Deng, X. Du, Z. Lu, Q. Duan, Y . He, S.-C. Huang, and J. Wu, “Laecips: large vision model assisted adaptive edge-cloud collaboration for iot-based perception system,”arXiv e-prints, pp. arXiv–2404, 2024

2024
[23]

Vavlm: Toward efficient edge-cloud video analytics with vision- language models,

Y . Zhang, H. Wang, Q. Bai, H. Liang, P. Zhu, G.-M. Muntean, and Q. Li, “Vavlm: Toward efficient edge-cloud video analytics with vision- language models,”IEEE Trans. Broadcast., 2025

2025
[24]

Dynrsl-vlm: Enhancing autonomous driving perception with dynamic resolution vision-language models,

X. Zhou, L. Shan, and X. Gui, “Dynrsl-vlm: Enhancing autonomous driving perception with dynamic resolution vision-language models,” arXiv preprint arXiv:2503.11265, 2025

2025 arXiv
[25]

Token communication in the era of large models: An information bottleneck- based approach,

H. Wei, W. Ni, W. Wang, W. Xu, D. Niyato, and P. Zhang, “Token communication in the era of large models: An information bottleneck- based approach,”IEEE Wireless Commun. Lett., vol. 15, pp. 186–190, 2026

2026
[26]

Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,

J. Li, D. Li, S. Savarese, and S. C. H. Hoi, “Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,” inProceedings of the 40th International Conference on Machine Learning. PMLR, 2023, pp. 19 730–19 742

2023
[27]

Llava-med: Training a large language- and-vision assistant for biomedicine in one day,

C. Li, C. Wong, S. Zhanget al., “Llava-med: Training a large language- and-vision assistant for biomedicine in one day,”Advances in Neural Information Processing Systems, vol. 36, pp. 28 541–28 564, 2023

2023
[28]

edgeVLM: Cloud-edge collaborative real-time VLM based on context transfer,

C. Qian, X. Yu, Z. Huang, D. Li, Q. Ma, F. Dang, X. Ding, G. Shang, and Z. Yang, “edgeVLM: Cloud-edge collaborative real-time VLM based on context transfer,” 2025

2025
[29]

Neely,Stochastic network optimization with application to commu- nication and queueing systems

M. Neely,Stochastic network optimization with application to commu- nication and queueing systems. Morgan & Claypool Publishers, 2010

2010
[30]

nuScenes: A multimodal dataset for autonomous driving,

H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Kr- ishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuScenes: A multimodal dataset for autonomous driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2020

2020

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Open- DriveVLA: Towards end-to-end autonomous driving with large vision language action model,

X. Zhou, X. Han, F. Yang, Y . Ma, V . Tresp, and A. Knoll, “Open- DriveVLA: Towards end-to-end autonomous driving with large vision language action model,” inProc. AAAI Conf. Artif. Intell. (AAAI), Mar 2026, pp. 13 782–13 790

2026

[2] [2]

VLM-MPC: Model predictive controller augmented vision language model for autonomous driving,

K. Long, H. Shi, J. Liu, C. Xiao, and X. Li, “VLM-MPC: Model predictive controller augmented vision language model for autonomous driving,”Transportation Research Part C: Emerging Technologies, vol. 183, p. 105487, 2026

2026

[3] [3]

A cloud–edge collabora- tive architecture for multimodal LLM-based advanced driver assistance systems in iot networks,

Y . Hu, D. Ye, J. Kang, M. Wu, and R. Yu, “A cloud–edge collabora- tive architecture for multimodal LLM-based advanced driver assistance systems in iot networks,”IEEE Internet Things J., vol. 12, no. 10, pp. 13 208–13 221, 2025

2025

[4] [4]

guided mllm reasoning: Enhancing mllm with knowledge and visual notes for visual question answering,

W. Fang, Q. Wu, J. Chen, and Y . Xue, “guided mllm reasoning: Enhancing mllm with knowledge and visual notes for visual question answering,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 19 597–19 607

2025

[5] [5]

EM-VLM4AD: Multi- frame, lightweight & efficient vision-language models for question answering in autonomous driving,

A. Gopalkrishnan, R. Greer, and M. Trivedi, “EM-VLM4AD: Multi- frame, lightweight & efficient vision-language models for question answering in autonomous driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) Workshops. IEEE, 2024, pp. 4552–4561

2024

[6] [6]

A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms,

W. Zhao, Y . Han, J. Tang, Z. Li, Y . Song, K. Wang, Z. Wang, and Y . You, “A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 19 814–19 824

2025

[7] [7]

Making large vision language models to be good few-shot learners,

F. Liu, W. Cai, J. Huo, C. Zhang, D. Chen, and J. Zhou, “Making large vision language models to be good few-shot learners,” inProc. AAAI Conf. Artif. Intell. (AAAI), 2025, pp. 5415–5423

2025

[8] [8]

Vision-language models for edge networks: A comprehensive survey,

A. Sharshar, L. U. Khan, W. Ullah, and M. Guizani, “Vision-language models for edge networks: A comprehensive survey,”IEEE Internet Things J., vol. 12, no. 16, pp. 32 701–32 724, 2025

2025