Mobile Traffic Prediction using LLMs with Efficient In-context Demonstration Selection

Akram Bin Sediq; Ali Afana; Han Zhang; Melike Erol-Kantarci

REVIEW 5 major objections 5 minor 44 references

Mobile Traffic Prediction using LLMs with Efficient In-context Demonstration Selection

T0 review · 5 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 The paper claims that two-step demonstration selection—similarity first, informativeness second—makes LLM-based 5G traffic prediction consistently more accurate than zero-shot and heuristic selection, without any fine-tuning.

desk verdict The two-step ICL selection is a real new variant for wireless traffic prediction and the cross-scenario generalization result is the paper's strongest asset, but the headline superiority claim over zero-shot is not yet cleanly established because the evaluation protocol leaks test-set information into the selection decisions. read the letter →

arxiv 2506.12074 v1 pith:J6TCRZT6 submitted 2025-06-05 cs.NI

Han Zhang , Akram Bin Sediq , Ali Afana , Melike Erol-Kantarci This is my paper

classification cs.NI

keywords mobiletrafficpredictionin-contextlearningdemonstrationselectionlargelanguagemodels5GnetworkskernelregressiontimeseriesforecastingR2score

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

This paper tries to establish that a pre-trained large language model can forecast real 5G downlink traffic accurately without fine-tuning, as long as the prompt includes one or two historical examples chosen by a specific two-step rule. The first step, effectiveness, picks the demonstration whose throughput series and incremental throughput series are closest to the test window; the second step, informativeness, keeps it only if the zero-shot model predicts that example poorly, on the grounds that an already-predictable example adds no new information. Both rules are derived from an analysis that treats in-context learning as kernel regression over a hidden Markov model of token generation. On a real 5G dataset spanning driving and static scenarios and downloading and video activities, the selected demonstrations give lower MAE and RMSE and higher R2-Scores than zero-shot prediction, constant demonstration selection, distance-only selection, and moving-average baselines. If the result holds, mobile operators could adapt forecasts to new traffic regimes by editing prompts instead of retraining or fine-tuning models.

What carries the argument

The load-bearing mechanism is the two-step selector: an effectiveness score $e_1+e_2$ computed from raw and incremental throughput distances, followed by an informativeness gate that compares the zero-shot error rate $E_r$ with a threshold. The theoretical object behind it is the kernel $K(x,x')=\langle \hat{P}_x^\top, \Sigma^{-1} \hat{P}_{x'}\rangle$ over the LLM's output-token probability vectors, together with the hidden Markov model $o_j\sim P(o_j=o|s_j=s)$ used to model token generation. The kernel-regression argument says ICL prediction error is bounded by $O(1-K(x_{\mathrm{te}},x_{\mathrm{ICL}}))$, so the first step maximizes similarity; the hidden-Markov and Bayesian argument says the benefit of a demonstration is the log-probability gap between a desired latent concept and the pre-trained concept, which the second step approximates by $E_r$. The batch-calibration step averages one-shot predictions over the selected demonstrations to reduce permutation sensitivity.

What would settle it

Compute the actual kernel $K(x_{\mathrm{te}},x_{\mathrm{D}})$ of Eq. (11) from the LLM's output-token probabilities for a set of 5G test windows and candidate demonstrations, and compare its ranking with the $e_1+e_2$ ranking. If the two rankings diverge substantially on samples with abrupt traffic changes or in sparse regions of the feature space, the effectiveness rule's distance proxy is falsified; a simpler check is whether prediction error tracks $e_1+e_2$ or the true kernel similarity when the two orderings conflict.

Watch

Extended reading notes

Core claim

The central claim is that ICL demonstration quality for LLM traffic prediction decomposes into two measurable properties. Effectiveness is quantified by $e_1+e_2$, where $e_1=\|\Gamma_{\mathrm{te}}-\Gamma_{\mathrm{D}}\|_2$ is the Euclidean distance between the historical throughput series and $e_2=\|\Delta\Gamma_{\mathrm{te}}-\Delta\Gamma_{\mathrm{D}}\|_2$ is the same distance for incremental (first-difference) series; the paper argues this is a computable proxy for maximizing the kernel similarity $K(x_{\mathrm{te}},x_{\mathrm{D}})$ between the LLM's output-token probability distributions, which in the kernel-regression view of ICL minimizes the error term $O(1-K(x_{\mathrm{te}},x_{\mathrm{ICL}}))$. Informativeness is quantified by the zero-shot prediction error rate $E_r=|\hat{y}_{\mathrm{D}}-y_{\mathrm{D}}|/y_{\mathrm{D}}$: a demonstration is used only when $E_r$ exceeds a threshold, because a pre-trained model that already predicts the demonstration correctly has nothing to learn from it. On real 5G data, one or two demonstrations selected this way, with one-shot predictions averaged as batch calibration, consistently outperform zero-shot and heuristic selection, lift R2-Scores above 0.65 across all tested scenarios, and keep R2 above 0.8 even when demonstrations come from a different activity scenario than the test data.

Load-bearing premise

The load-bearing premise is that Euclidean distance between raw and incremental throughput series faithfully measures the kernel similarity between the LLM's output-token probability distributions for those inputs; if that proxy fails on abrupt or non-smooth 5G traffic, the theoretical justification for the effectiveness step collapses, even though the heuristic might still work in practice.

Editorial extensions

If this is right

With the proposed selection, two demonstrations are enough to improve MAE, RMSE, and R2-Score over zero-shot prediction in every tested scenario, and adding more demonstrations can inject noise rather than accuracy.
The informativeness threshold creates a tunable cost-accuracy tradeoff: thresholds between 0.15 and 0.5 apply ICL to roughly half the test samples and give the best R2-Score, showing that many test inputs should be left in zero-shot mode.
Demonstrations drawn from a different application scenario still help, with R2 above 0.8 when test data come from a downloading-while-driving scenario, while an LSTM trained on the mismatched scenario degrades to near-zero R2.
A smaller LLM using selected demonstrations reaches approximately the accuracy of a larger LLM in zero-shot mode, suggesting careful demonstration choice can partly compensate for model capacity.
The selection method keeps test-time complexity linear in the number of candidate demonstrations, $O(n_d)$ for one demonstration, because the informativeness error rates can be precomputed offline.

Reading between the lines

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

Beyond the paper: the distance proxy is an empirical shortcut, and a natural test is to select demonstrations by the actual output-token kernel or by learned embeddings of traffic series; such a variant could be compared directly with the $e_1+e_2$ rule on the same 5G data.
Beyond the paper: the two-step logic, similarity first, then a check that the zero-shot model fails on the candidate, applies to any numeric forecasting task an LLM can read as text, such as energy load or edge-server demand, but the informativeness threshold would need per-task recalibration.
Beyond the paper: since the method treats the LLM as a black box and only edits the prompt, it could be combined with other prompt-level techniques like chain-of-thought or retrieval-augmented generation; the authors mention this only as future work.

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

Two things to know. First, the two-step selection rule—distance-based effectiveness plus zero-shot-error informativeness—is genuinely new for LLM-based wireless traffic prediction, and it is tested on real 5G data rather than synthetic smooth series. Second, the paper's load-bearing claim, that this selection beats zero-shot and simpler ICL baselines, is not yet cleanly established, because the evaluation protocol lets test-set information leak into the method's choices.

What the paper does well: the per-sample selection idea is sensible, the ablation shows both rules matter, and the comparison with LSTM is honest—they openly report that LSTM is more accurate when training and test come from the same scenario. The most credible positive finding is the cross-scenario generalization table: LLM predictions stay above R2=0.8 when demonstrations come from entirely different activities, while LSTM collapses. That is a real, practically useful result, and it does not depend on the contested cleaning steps.

Where it goes soft. The main empirical comparison is compromised by three choices. The informativeness threshold is selected after inspecting test-set R2 values (Table II), and the demonstration count M=2 is justified by "numerical results" in Section V.B. The zero-shot baseline is manually cleaned by removing hallucinated outliers before computing Table I and Fig. 3, with no pre-specified rule for what counts as a hallucination. There are no error bars and no code. Individually each is a minor flaw; together they make the headline gains unseparable from selection effects. The theoretical section has a real sign inconsistency: Eq. (18) says to minimize kernel similarity while the text and Eqs. (22)-(23) argue for maximizing it. That inconsistency is fixable, but it signals that the theory is more ornament than foundation. The unproven error decomposition in Eqs. (22)-(23) adds to that.

Verdict: a serious referee should engage with this, not desk-reject it. The cross-scenario result alone justifies referee time. But the authors should be required to run a validation split for the threshold, pre-register the hallucination-removal rule, and release code and data. The theory should either be corrected or downplayed. If the empirical claims survive a clean evaluation, the paper becomes a solid applied contribution; as it stands, it is a promising draft with an unproven headline.

I would bring this to a reading group and would cite the cross-scenario table in my own work, but I would not trust the headline numbers yet.

Referee Report

5 major / 5 minor

Summary. The paper proposes an LLM-based mobile traffic prediction framework that uses in-context learning (ICL) with a two-step demonstration selection strategy: an effectiveness rule that ranks candidate demonstrations by Euclidean distance between raw and incremental throughput series, and an informativeness rule that filters demonstrations based on the zero-shot prediction error rate relative to a threshold. The authors provide a theoretical analysis of ICL via hidden Markov models and kernel regression, and report experiments on a real 5G dataset with Phi-3 models, showing improved MAE, RMSE, and R2-Score over zero-shot prediction, distance-only ICL, constant ICL, and classical baselines such as moving averages, ARIMA, and LSTM.

Significance. If the reported gains survive a clean evaluation, the work is a valuable demonstration that a few carefully chosen ICL demonstrations can replace fine-tuning for mobile traffic prediction on realistic, non-smooth 5G traffic, with meaningful resource benefits in wireless networks. The paper uses a real-world 5G dataset, covers multiple application scenarios, and includes an ablation study and a comparison with LSTM-based methods, which strengthens its practical relevance. The attempt to ground the selection rules in kernel regression and Bayesian inference is also interesting. However, the central empirical claim is currently undermined by the evaluation protocol: the informativeness threshold and demonstration count are selected using test-set results, and zero-shot hallucinations are manually removed before computing metrics. These issues prevent the reported improvements from being interpreted as unbiased estimates of generalization, and the theoretical analysis contains an unresolved sign inconsistency.

major comments (5)

[Section V.B and Section V.C/Table II] The number of demonstrations M is fixed at two because "according to the numerical result, two demonstrations already can lead to a better performance" (Section V.B), and the informativeness threshold is selected by inspecting test-set R2 values in Table II; the headline results in Table I coincide with the threshold=0.35 column. Since no validation split or nested model-selection procedure is described, the reported improvements are not unbiased estimates of the method's generalization performance. Please rerun the evaluation with a validation-based protocol (e.g., selecting the threshold and M on a separate validation set) or report performance as a function of these hyperparameters with confidence intervals.
[Table I and Fig. 3] Table I states that "the hallucinations of zero-shot prediction have been manually removed" before computing metrics, and Fig. 3 removes "unstable outlier points" from the error visualization. Because no pre-specified, reproducible rule for identifying hallucinations is given, the comparison between the proposed method and the zero-shot baseline is not clean; manual cleaning can asymmetrically affect the baselines. Please apply an identical automated cleaning rule to all methods and report metrics both with and without cleaning.
[Eq. (18) and surrounding text] Eq. (18) instructs selecting demonstrations by minimizing the kernel similarity K(xte, xD), but the text immediately following Eq. (23) and the paragraph after it state that ICL reduces error when K is maximized and that "we maximize K(xtest, xICL)." This direct sign inconsistency is load-bearing for the effectiveness rule and must be resolved: either correct Eq. (18) to a maximization, or explain why the derived objective is a minimization despite the stated preference for similar demonstrations.
[Eqs. (22)-(23)] The error decomposition Er(ytest, ytest) = ϵnoise + ϵLLM and its ICL counterpart Er = ϵnoise + O(1−K(xtest, xICL)) are asserted without derivation. The link from the kernel-regression bound in Eq. (17) to this decomposition is not established, and the claim that ICL replaces ϵLLM with O(1−K) is not justified by the preceding equations. Please provide a derivation or explicitly present these expressions as modeling assumptions rather than proven results.
[Eqs. (19)-(21)] The kernel similarity between output-token probability distributions is approximated by the Euclidean distance between raw and incremental throughput series. No evidence is provided that this proxy is faithful for non-smooth 5G traffic; if the proxy fails, the theoretical grounding of the first selection step collapses, even if the heuristic happens to work. Please validate the proxy empirically, for example by correlating the proposed e1+e2 distance with the kernel similarity in Eq. (11) or with the downstream prediction error.

minor comments (5)

[Fig. 3 and Fig. 4(c)] The caption "(outliners removed)" in Fig. 3 should be "(outliers removed)", and "Imformativeness" in Fig. 4(c) should be "Informativeness".
[Around Eq. (15)] The sample size is denoted by lower-case m in Eq. (15) but by M elsewhere in the paper; please unify the notation.
[Section V.E] The model name is written inconsistently as "PHI-3-mini", "PHI-3-medium", and "Phi-3"; please use a single convention consistently.
[Reference [31]] The reference for the kernel-regression explanation of ICL is given as an arXiv preprint with a page number "p. 3"; please provide a complete citation or remove the page number.
[General] No error bars, standard deviations, or confidence intervals are reported for any of the numerical results, and no code or data availability statement is provided. Reporting variability across multiple runs would help assess the stability of the reported gains.

Circularity Check

1 steps flagged · score 4.0 of 10

Reported gains are partly a product of test-set hyperparameter selection; core selection heuristic is not definitionally circular.

fitted input called prediction [Section V.B (demonstration count), Section V.C and Tables I-II (error-rate threshold)]
"During the experiments, we select two ICL demonstrations since more demonstrations will lead to a higher computation cost and according to the numerical result, two demonstrations already can lead to a better performance and eliminate noise. ... According to the numerical results, setting the threshold between 0.5 and 0.15 is an ideal range to optimize the prediction accuracy and computational cost of ICL."

The headline result in Table I ('Selected two-demo prediction') is reported with M=2 demonstrations and with the threshold=0.35 column of Table II. Both hyperparameters are chosen from the test-set performance landscape: the demonstration count is selected 'according to the numerical result', and the threshold is selected because Table II shows it lands in the 'ideal range' on the same test set. The zero-shot baseline is not threshold-tuned, so the comparison is between a configuration optimized on the test set and a fixed baseline. The reported R2/MAE improvement is therefore partly a fitted quantity (test-set-selected hyperparameters) being presented as the method's predictive performance, rather than an out-of-sample prediction. No validation split is described.

full rationale

Most of the paper's derivation is not circular: the effectiveness and informativeness rules are heuristic selection criteria, not quantities defined in terms of the final prediction metric; the kernel-regression and HMM background is imported from external works [30]-[32]; and the one self-citation [14] is supporting context, not the load-bearing premise. The main circularity concern is in the empirical evaluation: the number of demonstrations and the informativeness threshold are chosen after inspecting test-set results (V.B, V.C, Tables I-II), so the reported superiority over zero-shot is partly an in-sample optimum. The manual removal of zero-shot hallucinations (V.B, Table I) and the sign/proxy inconsistency between Eq. (18) and Eq. (23) are further validity concerns, but they are correctness issues rather than definitional circularity. Consequently, the core selection method has independent content, but the headline 'prediction' result is partially fitted to the test set.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central method depends on a threshold and demonstration count chosen from the test data, and on several borrowed or ad hoc assumptions in the theoretical derivation. No new physical or architectural entities are introduced; effectiveness and informativeness are scoring rules, not entities.

free parameters (4)

Informativeness threshold for zero-shot error rate Er = 0.35 in main results; 0.5 to 0.15 declared ideal in Table II
The paper sweeps thresholds in Table II and selects the value giving the best R2/MAE; no a priori derivation is provided.
Number of ICL demonstrations M = 2
Chosen in Section V-B because two demonstrations already lead to better performance and eliminate noise; not derived from theory.
Historical window H = 5
Set in Section V-A without sensitivity analysis; the prediction depends on this input length.
Weights for e1 and e2 in the effectiveness score = 1:1
The sum e1+e2 in Eq. (21) is chosen without optimization; other weightings could change which demonstrations are selected.

assumptions (5)

domain assumption ICL behaves as implicit kernel regression with a kernel defined on output-token probability vectors.
Section IV-B imports this from [31] and builds the effectiveness rule on it without independent verification.
domain assumption LLM output generation can be modeled as a hidden Markov process with states and transition probabilities.
Section IV-B uses Eqs. (9)-(10) to motivate the informativeness analysis; this is a loose proxy, not a proven description of transformer inference.
ad hoc to paper The prior P(theta | x_D) is approximately uniform across the pre-trained concept theta_pre and the desired concept theta_0.
Section IV-C states this assumption to simplify Eq. (29) to Eq. (30); it is not justified beyond an appeal to lack of task-specific fine-tuning.
domain assumption The pre-trained output distribution is close to the task distribution, bounded by eta^2 epsilon_theta.
Eq. (16) assumes this closeness to derive the final bound; no evidence is given for the magnitude of epsilon_theta.
ad hoc to paper Kernel similarity between output-token distributions can be measured by Euclidean distance between raw and incremental throughput series.
Section IV-B replaces Eq. (18) with Eqs. (19)-(21) without proof, and Eq. (18) itself has a sign inconsistency.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mobile Traffic Prediction using LLMs with Efficient In-context Demonstration Selection." pith.science (2026). https://pith.science/paper/J6TCRZT6

@misc{pith2026250612074,
  author       = {Pith},
  title        = {Pith review of: Mobile Traffic Prediction using LLMs with Efficient In-context Demonstration Selection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J6TCRZT6}},
  note         = {Machine review of arXiv:2506.12074}
}

read the original abstract

Mobile traffic prediction is an important enabler for optimizing resource allocation and improving energy efficiency in mobile wireless networks. Building on the advanced contextual understanding and generative capabilities of large language models (LLMs), this work introduces a context-aware wireless traffic prediction framework powered by LLMs. To further enhance prediction accuracy, we leverage in-context learning (ICL) and develop a novel two-step demonstration selection strategy, optimizing the performance of LLM-based predictions. The initial step involves selecting ICL demonstrations using the effectiveness rule, followed by a second step that determines whether the chosen demonstrations should be utilized, based on the informativeness rule. We also provide an analytical framework for both informativeness and effectiveness rules. The effectiveness of the proposed framework is demonstrated with a real-world fifth-generation (5G) dataset with different application scenarios. According to the numerical results, the proposed framework shows lower mean squared error and higher R2-Scores compared to the zero-shot prediction method and other demonstration selection methods, such as constant ICL demonstration selection and distance-only-based ICL demonstration selection.

Figures

Figures reproduced from arXiv: 2506.12074 by the authors.

**Figure 2.** The workflow of the two-step ICL demonstration selection method. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** A comparison of traffic prediction with the proposed selected ICL demonstrations and zero-shot traffic prediction in [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: MAE and R2 -Score of mobile traffic prediction under different ICL settings. TABLE II: The ICL ratio, MAE, RMSE and R2 -Score under different predicted error rate threshold settings. Zero-shot prediction Threshold = 0.8 Threshold = 0.5 Threshold = 0.35 Threshold = 0.15…

**Figure 5.** Figure 5: The ICL ratio and the R2 -Score under different predicted error rate threshold settings. However, the superiority of the LSTM model comes at the cost of a large number of training computations and a rich training dataset. This will limit the use of LSTM models in some …

**Figure 6.** Figure 6: Traffic prediction with selected ICL demonstrations [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 27 canonical work pages

[1]

Heuristic algorithms for RIS-assisted wireless networks: Exploring heuristic-aided machine learning,

H. Zhou, M. Erol-Kantarci, Y . Liu, and H. V . Poor, “Heuristic algorithms for RIS-assisted wireless networks: Exploring heuristic-aided machine learning,” IEEE Wireless Communications , 2024

work page 2024
[2]

Intelligent reflecting surface assisted terahertz communications toward 6G,

Z. Chen, B. Ning, C. Han, Z. Tian, and S. Li, “Intelligent reflecting surface assisted terahertz communications toward 6G,” IEEE Wireless Communications, vol. 28, no. 6, pp. 110–117, 2021

work page 2021
[3]

The road towards 6G: A comprehensive survey,

W. Jiang, B. Han, M. A. Habibi, and H. D. Schotten, “The road towards 6G: A comprehensive survey,” IEEE Open Journal of the Communications Society , vol. 2, pp. 334–366, 2021

work page 2021
[4]

Transformer-based wireless traffic prediction and network optimization in O-RAN,

M. A. Habib, P. E. I. Rivera, Y . Ozcan, M. Elsayed, M. Bavand, R. Gaigalas, and M. Erol-Kantarci, “Transformer-based wireless traffic prediction and network optimization in O-RAN,” in 2024 IEEE Inter- national Conference on Communications Workshops (ICC Workshops) , 2024, pp. 1–6

work page 2024
[5]

Air traffic and usage predictions in avionic communications using attention based vaegan model,

H. Mukhtar, R. Schaub, and M. Erol-Kantarci, “Air traffic and usage predictions in avionic communications using attention based vaegan model,” in 2024 IEEE International Conference on Machine Learning for Communication and Networking (ICMLCN) , 2024, pp. 220–226

work page 2024
[6]

Weighted moving average forecast model based prediction service broker algorithm for cloud computing,

D. Kapgate, “Weighted moving average forecast model based prediction service broker algorithm for cloud computing,” International Journal of Computer Science and Mobile Computing , vol. 3, no. 2, pp. 71–79, 2014

work page 2014
[7]

Dual attention-based federated learning for wireless traffic prediction,

C. Zhang, S. Dang, B. Shihada, and M.-S. Alouini, “Dual attention-based federated learning for wireless traffic prediction,” in IEEE INFOCOM 2021-IEEE conference on computer communications . IEEE, 2021, pp. 1–10

work page 2021
[8]

RL meets multi-link operation in IEEE 802.11 be: Multi- headed recurrent soft-actor critic-based traffic allocation,

P. E. Iturria-Rivera, M. Chenier, B. Herscovici, B. Kantarci, and M. Erol- Kantarci, “RL meets multi-link operation in IEEE 802.11 be: Multi- headed recurrent soft-actor critic-based traffic allocation,” in ICC 2023- IEEE International Conference on Communications . IEEE, 2023, pp. 4001–4006

work page 2023

Show all 44 references

[9]

LLM-based intent processing and network optimization using attention-based hierarchical reinforcement learning,

M. A. Habib, P. E. I. Rivera, Y . Ozcan, M. Elsayed, M. Bavand, R. Gaigalas, and M. Erol-Kantarci, “LLM-based intent processing and network optimization using attention-based hierarchical reinforcement learning,” arXiv preprint arXiv:2406.06059 , 2024

2024 arXiv
[10]

Cbam: Convolutional block attention module,

S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 3–19

2018
[11]

Efficient multi-scale attention module with cross-spatial learning,

D. Ouyang, S. He, G. Zhang, M. Luo, H. Guo, J. Zhan, and Z. Huang, “Efficient multi-scale attention module with cross-spatial learning,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2023, pp. 1–5

2023
[12]

Phase shift compression for control signaling reduction in irs-aided wireless systems: Global attention and lightweight design,

X. Yu and D. Li, “Phase shift compression for control signaling reduction in irs-aided wireless systems: Global attention and lightweight design,” IEEE Transactions on Wireless Communications , vol. 23, no. 8, pp. 8528–8541, 2024

2024
[13]

One fits all: Power general time series analysis by pretrained LM,

T. Zhou, P. Niu, L. Sun, R. Jin et al. , “One fits all: Power general time series analysis by pretrained LM,” Advances in neural information processing systems, vol. 36, pp. 43 322–43 355, 2023

2023
[14]

Large language models in wireless application design: In-context learning-enhanced au- tomatic network intrusion detection,

H. Zhang, A. B. Sediq, A. Afana, and M. Erol-Kantarci, “Large language models in wireless application design: In-context learning-enhanced au- tomatic network intrusion detection,” arXiv preprint arXiv:2405.11002 , 2024

2024 arXiv
[15]

Large language models are zero-shot time series forecasters,

N. Gruver, M. Finzi, S. Qiu, and A. G. Wilson, “Large language models are zero-shot time series forecasters,” Advances in Neural Information Processing Systems, vol. 36, 2024

2024
[16]

Large language models can be easily distracted by irrelevant context,

F. Shi, X. Chen, K. Misra, N. Scales, D. Dohan, E. H. Chi, N. Sch ¨arli, and D. Zhou, “Large language models can be easily distracted by irrelevant context,” in International Conference on Machine Learning . PMLR, 2023, pp. 31 210–31 227

2023
[17]

Why larger language models do in-context learning differently?

Z. Shi, J. Wei, Z. Xu, and Y . Liang, “Why larger language models do in-context learning differently?” 2024

2024
[18]

Beyond throughput, the next generation: A 5G dataset with channel and context metrics,

D. Raca, D. Leahy, C. J. Sreenan, and J. J. Quinlan, “Beyond throughput, the next generation: A 5G dataset with channel and context metrics,” in Proceedings of the 11th ACM multimedia systems conference , 2020, pp. 303–308

2020
[19]

In-context learning with iterative demonstration selection,

C. Qin, A. Zhang, A. Dagar, and W. Ye, “In-context learning with iterative demonstration selection,” arXiv preprint arXiv:2310.09881 , 2023

2023 arXiv
[20]

Realtime mobile bandwidth and handoff predictions in 4G/5G networks,

L. Mei, J. Gou, Y . Cai, H. Cao, and Y . Liu, “Realtime mobile bandwidth and handoff predictions in 4G/5G networks,” Computer Networks , vol. 204, p. 108736, 2022

2022
[21]

A meta-learning scheme for adaptive short-term network traffic prediction,

Q. He, A. Moayyedi, G. D ´an, G. P. Koudouridis, and P. Tengkvist, “A meta-learning scheme for adaptive short-term network traffic prediction,” IEEE Journal on Selected Areas in Communications , vol. 38, no. 10, pp. 2271–2283, 2020

2020
[22]

Mobile data traffic prediction by exploiting time-evolving user mobility patterns,

F. Sun, P. Wang, J. Zhao, N. Xu, J. Zeng, J. Tao, K. Song, C. Deng, J. C. Lui, and X. Guan, “Mobile data traffic prediction by exploiting time-evolving user mobility patterns,” IEEE Transactions on mobile computing, vol. 21, no. 12, pp. 4456–4470, 2021

2021
[23]

Mobile traffic prediction from raw data using LSTM networks,

H. D. Trinh, L. Giupponi, and P. Dini, “Mobile traffic prediction from raw data using LSTM networks,” in 2018 IEEE 29th annual interna- tional symposium on personal, indoor and mobile radio communications (PIMRC). IEEE, 2018, pp. 1827–1832

2018
[24]

ST-Tran: Spatial-temporal transformer for cellular traffic prediction,

Q. Liu, J. Li, and Z. Lu, “ST-Tran: Spatial-temporal transformer for cellular traffic prediction,” IEEE Communications Letters , vol. 25, no. 10, pp. 3325–3329, 2021

2021
[25]

Performance analysis of network traffic predictors in the cloud,

B. L. Dalmazo, J. P. Vilela, and M. Curado, “Performance analysis of network traffic predictors in the cloud,” Journal of Network and Systems Management, vol. 25, pp. 290–320, 2017

2017
[26]

Network traffic prediction method based on au- toregressive integrated moving average and adaptive volterra filter,

Z. Tian and F. Li, “Network traffic prediction method based on au- toregressive integrated moving average and adaptive volterra filter,” International Journal of Communication Systems , vol. 34, no. 12, p. e4891, 2021

2021
[27]

Tempo: Prompt-based generative pre-trained transformer for time series forecasting,

D. Cao, F. Jia, S. O. Arik, T. Pfister, Y . Zheng, W. Ye, and Y . Liu, “Tempo: Prompt-based generative pre-trained transformer for time series forecasting,” arXiv preprint arXiv:2310.04948 , 2023

2023 arXiv
[28]

Llm4ts: Two-stage fine- tuning for time-series forecasting with pre-trained LLMs,

C. Chang, W.-C. Peng, and T.-F. Chen, “Llm4ts: Two-stage fine- tuning for time-series forecasting with pre-trained LLMs,”arXiv preprint arXiv:2308.08469, 2023

2023 arXiv
[29]

Self-refined generative foundation models for wireless traffic prediction,

C. Hu, H. Zhou, D. Wu, X. Chen, J. Yan, and X. Liu, “Self-refined generative foundation models for wireless traffic prediction,” arXiv preprint arXiv:2408.10390, 2024

2024
[30]

An explanation of in-context learning as implicit bayesian inference,

S. M. Xie, A. Raghunathan, P. Liang, and T. Ma, “An explanation of in-context learning as implicit bayesian inference,” arXiv preprint arXiv:2111.02080, 2021

2021 arXiv
[31]

In-context learning of large language models explained as kernel regression,

C. Han, Z. Wang, H. Zhao, and H. Ji, “In-context learning of large language models explained as kernel regression,” arXiv preprint arXiv:2305.12766, p. 3, 2023

2023 arXiv
[32]

Why can GPT learn in-context? language models implicitly perform gradient descent as meta-optimizers,

D. Dai, Y . Sun, L. Dong, Y . Hao, S. Ma, Z. Sui, and F. Wei, “Why can GPT learn in-context? language models implicitly perform gradient descent as meta-optimizers,” in ICLR 2023 Workshop on Mathematical and Empirical Understanding of F oundation Models

2023
[33]

Learning to retrieve prompts for in-context learning,

O. Rubin, J. Herzig, and J. Berant, “Learning to retrieve prompts for in-context learning,” arXiv preprint arXiv:2112.08633 , 2021

2021 arXiv
[34]

What makes good examples for visual in-context learning?

Y . Zhang, K. Zhou, and Z. Liu, “What makes good examples for visual in-context learning?” Advances in Neural Information Processing Systems, vol. 36, pp. 17 773–17 794, 2023

2023
[35]

Linkforecast: Cellular link bandwidth prediction in LTE networks,

C. Yue, R. Jin, K. Suh, Y . Qin, B. Wang, and W. Wei, “Linkforecast: Cellular link bandwidth prediction in LTE networks,” IEEE Transactions on Mobile Computing , vol. 17, no. 7, pp. 1582–1594, 2017

2017
[36]

Active example selection for in-context learning,

Y . Zhang, S. Feng, and C. Tan, “Active example selection for in-context learning,” arXiv preprint arXiv:2211.04486 , 2022

2022 arXiv
[37]

Llm-inference- bench: Inference benchmarking of large language models on ai acceler- ators,

K. T. Chitty-Venkata, S. Raskar, B. Kale, F. Ferdaus, A. Tanikanti, K. Raffenetti, V . Taylor, M. Emani, and V . Vishwanath, “Llm-inference- bench: Inference benchmarking of large language models on ai acceler- ators,” in SC24-W: Workshops of the International Conference for H...

2024
[38]

Latency-aware joint task offloading and energy control for cooperative mobile edge computing,

W. Fan, F. Xiao, Y . Pan, X. Chen, L. Han, and S. Yu, “Latency-aware joint task offloading and energy control for cooperative mobile edge computing,” IEEE Transactions on Services Computing , 2025

2025
[39]

Smoothquant: Accurate and efficient post-training quantization for large language models,

G. Xiao, J. Lin, M. Seznec, H. Wu, J. Demouth, and S. Han, “Smoothquant: Accurate and efficient post-training quantization for large language models,” in International Conference on Machine Learning . PMLR, 2023, pp. 38 087–38 099

2023
[40]

Why does in-context learning fail sometimes? evaluating in-context learning on open and closed questions,

X. Li, H. Tang, S. Chen, Z. Wang, R. Chen, and M. Abram, “Why does in-context learning fail sometimes? evaluating in-context learning on open and closed questions,” arXiv preprint arXiv:2407.02028 , 2024

2024 arXiv
[41]

What makes a good order of examples in in-context learning,

Q. Guo, L. Wang, Y . Wang, W. Ye, and S. Zhang, “What makes a good order of examples in in-context learning,” in Findings of the Association for Computational Linguistics ACL 2024 , 2024, pp. 14 892–14 904

2024
[42]

Rethinking the role of demonstrations: What makes in-context learning work?

S. Min, X. Lyu, A. Holtzman, M. Artetxe, M. Lewis, H. Hajishirzi, and L. Zettlemoyer, “Rethinking the role of demonstrations: What makes in-context learning work?” arXiv preprint arXiv:2202.12837 , 2022

2022 arXiv
[43]

Self-adaptive in-context learning: An information compression perspective for in-context example selec- tion and ordering,

Z. Wu, Y . Wang, J. Ye, and L. Kong, “Self-adaptive in-context learning: An information compression perspective for in-context example selec- tion and ordering,” in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers)...

2023
[44]

Phi- 3 technical report: A highly capable language model locally on your phone,

M. Abdin, S. A. Jacobs, A. A. Awan, J. Aneja, A. Awadallah, H. Awadalla, N. Bach, A. Bahree, A. Bakhtiari, H. Behl et al. , “Phi- 3 technical report: A highly capable language model locally on your phone,” arXiv preprint arXiv:2404.14219 , 2024

2024 arXiv

Pith tools

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

[1] [1]

Heuristic algorithms for RIS-assisted wireless networks: Exploring heuristic-aided machine learning,

H. Zhou, M. Erol-Kantarci, Y . Liu, and H. V . Poor, “Heuristic algorithms for RIS-assisted wireless networks: Exploring heuristic-aided machine learning,” IEEE Wireless Communications , 2024

work page 2024

[2] [2]

Intelligent reflecting surface assisted terahertz communications toward 6G,

Z. Chen, B. Ning, C. Han, Z. Tian, and S. Li, “Intelligent reflecting surface assisted terahertz communications toward 6G,” IEEE Wireless Communications, vol. 28, no. 6, pp. 110–117, 2021

work page 2021

[3] [3]

The road towards 6G: A comprehensive survey,

W. Jiang, B. Han, M. A. Habibi, and H. D. Schotten, “The road towards 6G: A comprehensive survey,” IEEE Open Journal of the Communications Society , vol. 2, pp. 334–366, 2021

work page 2021

[4] [4]

Transformer-based wireless traffic prediction and network optimization in O-RAN,

M. A. Habib, P. E. I. Rivera, Y . Ozcan, M. Elsayed, M. Bavand, R. Gaigalas, and M. Erol-Kantarci, “Transformer-based wireless traffic prediction and network optimization in O-RAN,” in 2024 IEEE Inter- national Conference on Communications Workshops (ICC Workshops) , 2024, pp. 1–6

work page 2024

[5] [5]

Air traffic and usage predictions in avionic communications using attention based vaegan model,

H. Mukhtar, R. Schaub, and M. Erol-Kantarci, “Air traffic and usage predictions in avionic communications using attention based vaegan model,” in 2024 IEEE International Conference on Machine Learning for Communication and Networking (ICMLCN) , 2024, pp. 220–226

work page 2024

[6] [6]

Weighted moving average forecast model based prediction service broker algorithm for cloud computing,

D. Kapgate, “Weighted moving average forecast model based prediction service broker algorithm for cloud computing,” International Journal of Computer Science and Mobile Computing , vol. 3, no. 2, pp. 71–79, 2014

work page 2014

[7] [7]

Dual attention-based federated learning for wireless traffic prediction,

C. Zhang, S. Dang, B. Shihada, and M.-S. Alouini, “Dual attention-based federated learning for wireless traffic prediction,” in IEEE INFOCOM 2021-IEEE conference on computer communications . IEEE, 2021, pp. 1–10

work page 2021

[8] [8]

RL meets multi-link operation in IEEE 802.11 be: Multi- headed recurrent soft-actor critic-based traffic allocation,

P. E. Iturria-Rivera, M. Chenier, B. Herscovici, B. Kantarci, and M. Erol- Kantarci, “RL meets multi-link operation in IEEE 802.11 be: Multi- headed recurrent soft-actor critic-based traffic allocation,” in ICC 2023- IEEE International Conference on Communications . IEEE, 2023, pp. 4001–4006

work page 2023