Attention-Enhanced Prompt Decision Transformers for UAV-Assisted Communications with AoI

Chi Lu; Jun Li; Shi Jin; Xiaoli Shi; Yiyang Ni; Zhe Wang

REVIEW 4 major objections 5 minor 12 references

Attention-Enhanced Prompt Decision Transformers for UAV-Assisted Communications with AoI

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read APDT, a Decision Transformer with attention over users, short prompts, and cost-to-go tokens, achieves 4.5x faster convergence and over 15% lower average AoI than conventional DRL, and 8% lower AoI than vanilla DT, while satisfying the…

desk verdict A plausible DT extension for UAV-AoI with a real protocol asymmetry problem that makes the headline DRL gains hard to verify. read the letter →

arxiv 2505.22170 v1 pith:KSHIAS4V submitted 2025-05-28 eess.SP cs.ITmath.IT

Chi Lu , Yiyang Ni , Zhe Wang , Xiaoli Shi , Jun Li , Shi Jin This is my paper

classification eess.SPcs.ITmath.IT

keywords UAV-assistedcommunicationsageofinformationDecisionTransformerdeepreinforcementlearningpathplanninguserschedulingattentionmechanismprompt

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

The paper claims that a Decision Transformer (DT) extended with an attention layer over users, short trajectory prompts, and a cost-to-go token learns UAV flight and user-scheduling policies that minimize average age of information (AoI) while respecting a long-term energy budget. In simulation, the proposed APDT converges 4.5 times faster and cuts average AoI by over 15% compared to conventional deep reinforcement learning (DRL), and beats conventional DT by about 8% average AoI with twice the convergence speed, under dynamic user arrivals and departures. The significance is a path toward offline-pretrained sequence models that adapt to new wireless scenarios without retraining from scratch.

What carries the argument

The central object is the attention-enhanced prompt Decision Transformer (APDT), a sequence-modeling policy trained by supervised action prediction. The machinery consists of three parts: a set-attention aggregation defined in Eq. (6) that compresses any number of user states into a fixed-size representation using the UAV position as the query; a prompt mechanism that prepends a K-length demonstration segment sampled from an online buffer to condition the autoregressive model on the current scenario; and a token-assisted method that adds a cost-to-go token C(t) alongside the reward-to-go R(t) so the model represents the cumulative energy budget as part of the sequence. These components are trained end-to-end by minimizing the mean squared action-prediction loss (Eq. (5)) over offline expert trajectories, and the same architecture is deployed online with an updating prompt buffer.

What would settle it

An ablation that replaces APDT's attention aggregation with zero-padding but keeps the prompt and cost-to-go tokens, compared on the same dynamic-user scenarios, would settle whether the attention mechanism is the source of the reported AoI and convergence gains over vanilla DT.

Watch

Extended reading notes

Core claim

The paper's central claim is that three enhancements fix three known flaws of DT in this setting. Zero-padding for variable user counts is replaced by a multi-head attention operation, with the UAV's position projected as the query and each user's position-AoI pair as keys and values, producing a fixed-size representation without padding. A short demonstration segment of K steps is prepended to the input sequence as a prompt, letting the pretrained model adapt to a new scenario without gradient updates. A cost-to-go token C(t) is inserted alongside the reward-to-go, so the transformer learns a mapping from future cumulative energy consumption to actions and can meet a long-term energy limit. The combined framework is pretrained offline on expert trajectories collected in fixed-user environments, then deployed online with a small experience buffer that supplies fresh prompts. Simulation results report a 4.5x faster convergence and over 15% lower average AoI than conventional DRL, and about 8% lower AoI with twice the convergence speed than conventional DT.

Load-bearing premise

The load-bearing premise is that a fixed-size attention aggregation of a variable number of user states preserves enough information for near-optimal scheduling decisions.

Editorial extensions

If this is right

An offline-pretrained APDT can be deployed directly in UAV-assisted IoT scenarios with changing user counts, avoiding per-scenario retraining from scratch.
The cost-to-go token provides a template for embedding additive long-term constraints, such as total energy, into transformer-based decision models.
The prompt mechanism allows few-shot adaptation from demonstrations as short as K steps, reducing the need to collect large expert datasets for each new environment.
Because the performance gap over DT widens as user density increases, the attention aggregation is most beneficial in congested, highly dynamic service areas.
The framework can be pretrained on a handful of fixed-user environments and still serve a range of user densities between and beyond those settings.

Reading between the lines

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

An ablation that swaps the attention aggregation for zero-padding while keeping the prompt and cost-to-go tokens would test whether attention is the active ingredient behind the reported AoI gains or whether the prompt and token additions carry most of the improvement.
Because the expert dataset is generated by a converged TD3-Lag agent, APDT's ceiling is tied to that expert's quality; pretraining on self-generated or mixed-expert data would decouple the transformer's generalization from the expert's bias.
The cost-to-go token idea naturally fits additive cumulative constraints like total energy; applying the same mechanism to peak-power or per-slot latency limits would need different machinery, so the method's scope is narrower than general constrained sequence-modeling RL.

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

Here's my read on arXiv:2505.22170. The paper combines three known ideas—attention-based state aggregation, prompt-based adaptation, and cost-to-go tokens—into a Decision Transformer for UAV-assisted AoI minimization. The specific combination is new for this problem, and the attention aggregation is a sensible way to handle the variable user count without zero-padding. The energy-token approach is also a reasonable way to fold a long-term constraint into sequence modeling. The problem formulation is clear, and Algorithm 1 spells out the two-phase procedure well.

The soft spots are mostly experimental. The headline claim of 4.5x faster convergence and >15% lower AoI than DRL looks fragile because the protocol is unbalanced. In Algorithm 1, APDT actively adapts during deployment by filling an online prompt buffer, while Section V only says other algorithms were pre-trained before evaluation. If SAC-Lag/TD3-Lag/IPO are frozen after pretraining, the comparison measures adaptation protocol, not architecture. The paper should say explicitly whether baselines were allowed to continue learning or fine-tune in the ρ=20 scenario. Without that, the DRL comparison is not convincing.

Also missing: error bars, ablations (e.g., with vs. without attention, prompt, or energy token), and any plot showing the energy constraint is actually satisfied. The attention equation (6) is under-specified—it shows softmax of a scalar but never defines the aggregation across users; a sum or mean is implied but not written. Minor issue: the offline expert data is generated by TD3-Lag, itself a baseline, which is fine but should be discussed as a potential source of bias.

The DT comparison is more credible—DT also fine-tunes online, so the 8% AoI improvement is a fairer head-to-head. But even that would benefit from confidence intervals.

Overall, the core idea is promising and the paper is clearly written. It deserves a serious referee; a revised version with a balanced baseline protocol, ablations, and a properly specified attention mechanism could be a solid contribution to the UAV comms community. For now, I'd treat the quantitative claims as provisional.

Recommendation: send to peer review, but expect heavy revision.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes an Attention-Enhanced Prompt Decision Transformer (APDT) for UAV-assisted communications in which the number of terrestrial users varies over time. The method augments a Decision Transformer with an attention mechanism over user states (to avoid zero-padding), a prompt-based adaptation module built from short demonstration segments, and a token-assisted cost-to-go mechanism intended to enforce the UAV's long-term energy constraint. The problem is formulated as a CMDP, the model is pre-trained offline on datasets with 11, 13, and 15 fixed users collected by a TD3-Lag agent, and evaluated in new dynamic-user scenarios with average density ρ=20 and across ρ=15–25. The paper reports that APDT converges 4.5 times faster and reduces average AoI by over 15% compared to conventional DRL algorithms, and achieves over 8% lower average AoI with twice faster convergence compared to conventional DT.

Significance. If the reported gains were robust, the paper would make a useful contribution to AoI-aware UAV networking by showing that offline pretraining with attention-based variable-size state handling and prompt-based adaptation can transfer to dynamic user scenarios more efficiently than standard DRL. The CMDP formulation is clear, and the idea of using a cost-to-go token for constraint satisfaction is an interesting and potentially valuable departure from standard DT. The authors are transparent about using a TD3-Lag agent to generate the offline expert data, which is a strength in reproducibility. That said, the empirical evidence as presented does not yet support the headline numbers: there are no error bars, no ablations, no energy-consumption curves, and the comparison protocol between APDT and the DRL baselines is ambiguous.

major comments (4)

[Section V / Algorithm 1] The comparison with the DRL baselines is not balanced. Algorithm 1 allows APDT and DT to adapt during online deployment (APDT fills a prompt buffer; DT performs real-time fine-tuning using TD3-Lag samples), while Section V states only that SAC-Lag, TD3-Lag, and IPO were “pre-trained in these scenarios before evaluation.” If those baselines are evaluated as frozen policies, the reported 4.5x/15% advantage measures the benefit of in-context adaptation rather than the proposed architecture, and a “convergence speed” comparison is undefined for frozen agents. The authors should either let all baselines continue training under the same online interaction budget or explicitly frame the comparison as offline-pretrained-plus-adaptation versus frozen pretrained DRL, and adjust the claims accordingly.
[Eq. (6), Section IV] The attention equation does not implement the described mechanism. For a fixed user u, the softmax argument is (W_Q q(t))^T (W_K s_u(t)) / sqrt(d_k), a scalar; the softmax of a scalar is 1, and no summation or aggregation over u appears. Thus Eq. (6) cannot compress the variable-size set of users into a fixed representation, which is the central purpose of the proposed attention module. The authors need to specify the intended multi-head attention over all u in K(t), with proper query/key/value dimensions and an explicit aggregation formula, or state the exact operation used in the implementation.
[Section IV / Algorithm 1 / Fig. 3] The energy-constraint mechanism is underspecified and its satisfaction is not evidenced. The paper introduces C(t) as cost-to-go but does not define how C(1) is initialized in online deployment, how the target cost is chosen, or why the supervised mapping from C(t) to actions guarantees feasibility of constraint (4f). The only evidence is the Fig. 3 caption “energy constraint satisfied after convergence, except for DT,” with no energy-consumption curve, cumulative-energy plot, or constraint-violation metric shown. Please specify the constraint-handling mechanism and report the evolution of cumulative energy or constraint violations over episodes.
[Section V / Figures 3 and 4] The quantitative claims are not statistically supported. Figures 3 and 4 show single curves with no error bars or number of random seeds, “convergence” is never defined, and values such as 4.5x/2x faster or 15%/8% lower AoI are read off curves without a measurement protocol. The authors should provide mean values with standard deviations over multiple seeds, a precise convergence criterion (for example, episodes needed to reach within X% of the final performance), and preferably a table with per-density results so that the claimed margins can be verified.

minor comments (5)

[Abstract / Section I] The manuscript contains several typos and grammatical errors, including “Howerver” in Section I, “convenional” in the Abstract, “the calculated by” in Section II-A, and “mimimize” in Section I. A careful proofread is needed.
[Section III] The “linear method” used to map the discrete user-association action ξ(t) into a continuous value is not defined; please describe the encoding or the intended relaxation.
[Section II-B] The AoI update in Eq. (1) resets the user’s AoI to 1 in the time slot immediately after service, which does not explicitly account for the service duration δ_h(t) or the packet transmission time; please clarify when AoI is measured and how partial service is handled.
[Figures 3 and 4] The caption of Fig. 3 says “energy constraint satisfied after convergence, except for DT,” but the figure does not show energy; adding an energy panel or a separate constraint-violation plot would make the claim verifiable.
[References] The original Decision Transformer paper is not cited; the manuscript cites only the wireless-specific DT paper [8] when introducing DT. Please cite the original method (Chen et al., 2021) as well.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity; the paper's claims are empirical and self-contained, with only a minor non-load-bearing self-citation in the motivation.

full rationale

The paper's derivation chain is a standard supervised-learning pipeline: offline expert trajectories are collected from a converged TD3-Lag agent, APDT is trained to imitate expert actions conditioned on return-to-go and cost-to-go tokens, and the resulting policy is evaluated in new dynamic-user scenarios against DT, SAC-Lag, TD3-Lag, and IPO. No equation in the paper reduces a claimed prediction to a fitted input. The attention mechanism in Eq. (6) is a declared construction for variable-size user states, not a result derived from itself. The energy-constraint token method is transparent about its source: Section IV states that the APDT model is pretrained on a dataset that satisfies the long-term energy consumption constraint, so the later empirical statement that all methods satisfy the energy threshold after convergence is a learned behavior inherited from feasible demonstrations, not an identity. This is an overclaim about ensuring the constraint, but it is not a circular reduction. The only self-citation is [8], by overlapping authors (Jun Li, Shi Jin, Zhe Wang), used to introduce DT and its zero-padding limitation; since the paper also runs DT as a baseline in the same simulations, the self-citation is not load-bearing. The potentially unbalanced comparison with conventional DRL baselines (which may be frozen after pretraining while APDT adapts via online prompts) is an experimental-protocol concern, not a circularity. The absence of a theoretical analysis for the attention aggregation is a missing-support concern, not circularity. Accordingly, I find no specific circular step, and the mild self-citation warrants only the low end of the scale.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central improvements are empirical and depend on the simulation environment and the quality of the offline dataset. The method does not introduce new physical entities, and the only explicit hand-chosen hyperparameters are the prompt length and buffer size. Most assumptions are inherited from prior work on Decision Transformers and wireless channel modeling.

free parameters (2)

prompt length K = 5
Hand-chosen hyperparameter in the online buffer and prompt; no sensitivity analysis.
online buffer size O = 500
Hand-chosen hyperparameter for prompt sampling; no sensitivity analysis.

assumptions (3)

domain assumption The Gauss-Markov mobility model and the channel model from reference [10] accurately model the UAV-user communication environment.
All simulated results rest on these model choices; if they are unrealistic, the quantitative gains may not transfer.
domain assumption The offline dataset collected from a converged TD3-Lag agent in fixed-user scenarios is expert and representative enough for pre-training APDT.
The quality and coverage of the expert data determine the ceiling of the learned policy; the paper provides no analysis of dataset coverage or baselines.
domain assumption A Decision Transformer can learn a good mapping from (return, cost-to-go, state) sequences to actions from offline data.
This is the core ML assumption of the method, inherited from reference [8]; it is not proven here.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Attention-Enhanced Prompt Decision Transformers for UAV-Assisted Communications with AoI." pith.science (2026). https://pith.science/paper/KSHIAS4V

@misc{pith2026250522170,
  author       = {Pith},
  title        = {Pith review of: Attention-Enhanced Prompt Decision Transformers for UAV-Assisted Communications with AoI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KSHIAS4V}},
  note         = {Machine review of arXiv:2505.22170}
}

abstract

Decision Transformer (DT) has recently demonstrated strong generalizability in dynamic resource allocation within unmanned aerial vehicle (UAV) networks, compared to conventional deep reinforcement learning (DRL). However, its performance is hindered due to zero-padding for varying state dimensions, inability to manage long-term energy constraint, and challenges in acquiring expert samples for few-shot fine-tuning in new scenarios. To overcome these limitations, we propose an attention-enhanced prompt Decision Transformer (APDT) framework to optimize trajectory planning and user scheduling, aiming to minimize the average age of information (AoI) under long-term energy constraint in UAV-assisted Internet of Things (IoT) networks. Specifically, we enhance the convenional DT framework by incorporating an attention mechanism to accommodate varying numbers of terrestrial users, introducing a prompt mechanism based on short trajectory demonstrations for rapid adaptation to new scenarios, and designing a token-assisted method to address the UAV's long-term energy constraint. The APDT framework is first pre-trained on offline datasets and then efficiently generalized to new scenarios. Simulations demonstrate that APDT achieves twice faster in terms of convergence rate and reduces average AoI by $8\%$ compared to conventional DT.

Figures

Figures reproduced from arXiv: 2505.22170 by the authors.

**Figure 1.** UAV-assisted communications scenario. II. SYSTEM MODEL In this letter, we consider uplink communication from terrestrial mobile users assisted by an UAV, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** An overview of the APDT framework. Initially, an offline dataset [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Average AoI performance of different algorithms in the new scenario, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Average AoI performance of different algorithms across different [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages

[1]

Personalized federated learning with differential privacy and convergence guarantee,

K. Wei, J. Li, C. Ma, M. Ding, W. Chen, J. Wu, M. Tao, and H. V . Poor, “Personalized federated learning with differential privacy and convergence guarantee,”IEEE Trans. Inf. F orensics Secur ., vol. 18, pp. 4488–4503, 2023

work page 2023
[2]

Trusted AI in multiagent systems: An overview of privacy and security for distributed learning,

C. Ma, J. Li, K. Wei, B. Liu, M. Ding, L. Yuan, Z. Han, and H. Vincent Poor, “Trusted AI in multiagent systems: An overview of privacy and security for distributed learning,”Proc. IEEE, vol. 111, no. 9, pp. 1097–1132, Sep. 2023

work page 2023
[3]

Age of information: An introduction and survey,

R. D. Yates, Y . Sun, D. R. Brown, S. K. Kaul, E. Modiano, and S. Ulukus, “Age of information: An introduction and survey,”IEEE J. Sel. Areas Commun., vol. 39, no. 5, pp. 1183–1210, May 2021

2021
[4]

Decentralized federated reinforcement learning for user-centric dynamic TFDD con- trol,

Z. Yin, Z. Wang, J. Li, M. Ding, W. Chen, and S. Jin, “Decentralized federated reinforcement learning for user-centric dynamic TFDD con- trol,”IEEE J. Sel. Top. Signal Process, vol. 17, no. 1, pp. 40–53, Jan. 2023

work page 2023
[5]

AoI-minimal trajectory planning and data collection in UA V-assisted wireless powered IoT networks,

H. Hu, K. Xiong, G. Qu, Q. Ni, P. Fan, and K. B. Letaief, “AoI-minimal trajectory planning and data collection in UA V-assisted wireless powered IoT networks,”IEEE Internet Things J., vol. 8, no. 2, pp. 1211–1223, Jan. 2021

work page 2021
[6]

Average AoI minimization in UA V-assisted data collection with RF wireless power transfer: A deep reinforcement learning scheme,

L. Liu, K. Xiong, J. Cao, Y . Lu, P. Fan, and K. B. Letaief, “Average AoI minimization in UA V-assisted data collection with RF wireless power transfer: A deep reinforcement learning scheme,”IEEE Internet Things J., vol. 9, no. 7, pp. 5216–5228, Apr. 2022

work page 2022
[7]

Age of Information Aware Trajectory Planning of UA Vs in Intelligent Transportation Systems: A Deep Learning Approach,

M. Samir, C. Assi, S. Sharafeddine, D. Ebrahimi, and A. Ghrayeb, “Age of Information Aware Trajectory Planning of UA Vs in Intelligent Transportation Systems: A Deep Learning Approach,”IEEE Trans. V eh. Technol., vol. 69, no. 11, pp. 12 382–12 395, Nov. 2020

work page 2020
[8]

Decision Transformers for Wireless Communications: A New Paradigm of Resource Management,

J. Zhang, J. Li, Z. Wang, L. Shi, S. Jin, W. Chen, and H. V . Poor, “Decision Transformers for Wireless Communications: A New Paradigm of Resource Management,”IEEE Wireless Commun., pp. 1–7, 2025

work page 2025

Show all 12 references

[9]

Path planning for UA V-mounted mobile edge computing with deep reinforcement learning,

Q. Liu, L. Shi, L. Sun, J. Li, M. Ding, and F. Shu, “Path planning for UA V-mounted mobile edge computing with deep reinforcement learning,”IEEE Trans. V eh. Technol., vol. 69, no. 5, pp. 5723–5728, May 2020

2020
[10]

Opponent modeling based dynamic resource trading for UA V-assisted edge com- puting,

J. Bai, Z. Wang, J. Li, L. Shi, J. Zhang, K. Wei, and H. He, “Opponent modeling based dynamic resource trading for UA V-assisted edge com- puting,” inProc. IEEE 98th V eh. Technol. Conf. (VTC2023-Fall), Oct. 2023, pp. 1–6

2023
[11]

Benchmarking safe exploration in deep reinforcement learning

A. Ray, J. Achiam, and D. Amodei, “Benchmarking safe exploration in deep reinforcement learning.” OpenAI, 2019

2019
[12]

Ipo: Interior-Point Policy Optimization under Constraints,

Y . Liu, J. Ding, and X. Liu, “Ipo: Interior-Point Policy Optimization under Constraints,”Proc. AAAI Conf. Artif. Intell., vol. 34, no. 04, pp. 4940–4947, 2020

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Personalized federated learning with differential privacy and convergence guarantee,

K. Wei, J. Li, C. Ma, M. Ding, W. Chen, J. Wu, M. Tao, and H. V . Poor, “Personalized federated learning with differential privacy and convergence guarantee,”IEEE Trans. Inf. F orensics Secur ., vol. 18, pp. 4488–4503, 2023

work page 2023

[2] [2]

Trusted AI in multiagent systems: An overview of privacy and security for distributed learning,

C. Ma, J. Li, K. Wei, B. Liu, M. Ding, L. Yuan, Z. Han, and H. Vincent Poor, “Trusted AI in multiagent systems: An overview of privacy and security for distributed learning,”Proc. IEEE, vol. 111, no. 9, pp. 1097–1132, Sep. 2023

work page 2023

[3] [3]

Age of information: An introduction and survey,

R. D. Yates, Y . Sun, D. R. Brown, S. K. Kaul, E. Modiano, and S. Ulukus, “Age of information: An introduction and survey,”IEEE J. Sel. Areas Commun., vol. 39, no. 5, pp. 1183–1210, May 2021

2021

[4] [4]

Decentralized federated reinforcement learning for user-centric dynamic TFDD con- trol,

Z. Yin, Z. Wang, J. Li, M. Ding, W. Chen, and S. Jin, “Decentralized federated reinforcement learning for user-centric dynamic TFDD con- trol,”IEEE J. Sel. Top. Signal Process, vol. 17, no. 1, pp. 40–53, Jan. 2023

work page 2023

[5] [5]

AoI-minimal trajectory planning and data collection in UA V-assisted wireless powered IoT networks,

H. Hu, K. Xiong, G. Qu, Q. Ni, P. Fan, and K. B. Letaief, “AoI-minimal trajectory planning and data collection in UA V-assisted wireless powered IoT networks,”IEEE Internet Things J., vol. 8, no. 2, pp. 1211–1223, Jan. 2021

work page 2021

[6] [6]

Average AoI minimization in UA V-assisted data collection with RF wireless power transfer: A deep reinforcement learning scheme,

L. Liu, K. Xiong, J. Cao, Y . Lu, P. Fan, and K. B. Letaief, “Average AoI minimization in UA V-assisted data collection with RF wireless power transfer: A deep reinforcement learning scheme,”IEEE Internet Things J., vol. 9, no. 7, pp. 5216–5228, Apr. 2022

work page 2022

[7] [7]

Age of Information Aware Trajectory Planning of UA Vs in Intelligent Transportation Systems: A Deep Learning Approach,

M. Samir, C. Assi, S. Sharafeddine, D. Ebrahimi, and A. Ghrayeb, “Age of Information Aware Trajectory Planning of UA Vs in Intelligent Transportation Systems: A Deep Learning Approach,”IEEE Trans. V eh. Technol., vol. 69, no. 11, pp. 12 382–12 395, Nov. 2020

work page 2020

[8] [8]

Decision Transformers for Wireless Communications: A New Paradigm of Resource Management,

J. Zhang, J. Li, Z. Wang, L. Shi, S. Jin, W. Chen, and H. V . Poor, “Decision Transformers for Wireless Communications: A New Paradigm of Resource Management,”IEEE Wireless Commun., pp. 1–7, 2025

work page 2025