PERA: A Perceive-Reason-Act Interface Bridging Sensing, Cognitive Reasoning, and Trustworthy Agentic Response for 6G

Lajos Hanzo; Melike Erol-Kantarci; Mohammad Farzanullah

REVIEW 3 major objections 6 minor 16 references

A single language-model interface that reads wireless telemetry can replace task-specific edge models for physical-layer diagnostics and beam control, outperforming them on unseen urban scenarios while emitting human-readable causal rationa

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 19:34 UTC pith:EMA2QVON

load-bearing objection A competent transfer of the LLaVA-style projector recipe to wireless CSI, with a plausible but under-evidenced case study; the explainability claim is currently asserted, not shown. the 3 major comments →

arxiv 2607.16896 v1 pith:EMA2QVON submitted 2026-07-18 cs.NI

PERA: A Perceive-Reason-Act Interface Bridging Sensing, Cognitive Reasoning, and Trustworthy Agentic Response for 6G

Mohammad Farzanullah , Melike Erol-Kantarci , Lajos Hanzo This is my paper

classification cs.NI

keywords 6G networksagentic AIlarge language modelswireless foundation modelsmulti-modal instruction tuningexplainable AIbeam predictionDeepMIMO

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes PERA, a Perceive-Reason-Act interface that pairs a large wireless AI model (sensory) with a large language model (cognitive) through a lightweight learnable projector. It argues that this single generative interface can replace the fragmented, task-specific models that currently run at the network edge. On the DeepMIMO dataset, the approach classifies line-of-sight versus non-line-of-sight links and predicts beam groups on six previously unseen cities, with F1 gains of +7.73% and +9.90% over raw-CSI CNN baselines respectively. The system also produces human-readable rationales that ground each decision in the channel's energy distribution, which the authors argue is the key to operator trust in 6G automation.

Core claim

PERA's central claim is that grounding an LLM in wireless telemetry, rather than adapting the LLM to the wireless domain or vice versa, yields a unified edge intelligence that is more accurate and more explainable than task-specific discriminative models. The paper demonstrates this by tokenizing raw CSI with a frozen large wireless model, projecting the resulting embeddings into the LLM's space with a two-layer MLP bridge, and then fine-tuning LoRA adapters with a weighted multi-task objective that combines language modeling, link-state classification, and beam-group prediction. On novel DeepMIMO cities, this interface outperforms a raw-CSI CNN by 7.73% for LoS/NLoS F1 and by 31.17% for bea

What carries the argument

The load-bearing mechanism is the 'modality bridge' (Layer 2), a two-layer MLP projector that maps latent vectors from a frozen wireless perceptual encoder into the embedding space of a frozen LLM, yielding 'network pseudo-tokens' that the LLM reads as text. This is trained in two stages: first a representation-alignment stage that teaches the bridge to map channel states to descriptive captions, then a goal-oriented instruction-tuning stage that trains the bridge together with LoRA adapters in the LLM, augmented by auxiliary LoS and beam-group heads whose losses are weighted 0.3:1.0:0.5 against the language loss to keep outputs physically faithful.

Load-bearing premise

The language model's decisions and explanations are actually grounded in the wireless tokens rather than generated from language priors or hallucination, and this grounding persists in every new deployment environment.

What would settle it

Take the trained PERA model, fix the text prompt, and randomly permute or zero-out the network pseudo-tokens at inference. If the predicted beam group and the generated rationale remain largely unchanged, the system is not reading the channel telemetry and the central explainability claim collapses. Alternatively, test on real measured channels from a city not in the DeepMIMO training set: if the F1 gains over the raw-CSI CNN disappear, the grounding does not transfer.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

A single PERA-style core could replace hundreds of task-specific edge diagnostic models, because new tasks or environments are handled by swapping LoRA adapters or editing prompts rather than retraining.
Because the interface outputs control parameters alongside natural-language rationales, it provides operator-visible audit trails for physical-layer decisions, which is a step toward trusted autonomous network control.
The same three-tier architecture is layer-agnostic: routing, slicing, and intent-negotiation tasks could in principle reuse the same perceptual encoder and LLM core with different heads or instructions.
Zero-shot adaptation to previously unseen environments is demonstrated: the model trained on five cities classifies and predicts beams in six new cities without gradient updates, suggesting that prompt-based adaptation at inference time is possible for novel scenarios.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the grounding holds, the PERA recipe generalizes beyond wireless: any high-dimensional sensor stream (radar, optical, audio) could be projected into an LLM's embedding space and reasoned over with the same pseudo-token mechanism, turning LLMs into a universal reasoning layer over physical modalities.
The paper does not ablate the auxiliary physical-fidelity heads, so the exact source of the accuracy gain — the LLM's reasoning versus the shared LWM features — remains untested. A version without the heads, or with a frozen bridge, would isolate the contribution of language modeling to the F1 improvements.
The trustworthiness claim hinges on explanation faithfulness, which the paper does not quantitatively measure. A natural next step is to corrupt or mask the wireless pseudo-tokens while holding the text prompt fixed; if the output beam or rationale does not change, the rationales are not causally grounded in the channel state.
Since the experiments use the simulated DeepMIMO dataset, the decisive test for real 6G deployment is whether the demonstrated grounding survives real over-the-air channel measurements, hardware impairments, and distribution drift beyond the six test cities.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A competent transfer of the LLaVA-style projector recipe to wireless CSI, with a plausible but under-evidenced case study; the explainability claim is currently asserted, not shown.

read the letter

Quick take: this is a legitimate architectural transfer worth a serious look, but the experimental section is not yet solid enough to support the headline claims—especially the grounding/trust story.

What is actually new: applying the well-known frozen encoder + trainable projector + LoRA-adapted LLM recipe from visual instruction tuning to physical-layer wireless telemetry, and adding auxiliary heads for link state and beam groups to keep the LLM honest. The framing of 'network pseudo-tokens' is more presentation than substance, but the two-stage curriculum with the physical-fidelity heads is a reasonable engineering choice. I have not seen this exact combo on DeepMIMO-style CSI with a modern LLM.

What it does well: the architecture is clearly described, the literature review is competent and not padded, and the case study is a real generalization check—trained on five cities, tested on six unseen ones. That is more than many wireless-LLM papers do.

Where it falls short, in order of importance:

1. The explainability claim is the differentiator and it is not evaluated at all. The only evidence is one hand-picked example in Fig. 5, with no check that the 'peak @ 122' or 'valley @ 0' actually correspond to real positions in the tokenized channel. No faithfulness metric, no protocol, no comparison of the auxiliary-weighted objective (0.3:1.0:0.5) against a version without the auxiliary heads or without the language loss. I agree with the stress-test note: if those rationales are post-hoc smooth-talk, the trust claim collapses. This needs an ablation and a quantitative faithfulness measure before I believe it.

2. The empirical numbers are under-specified. No error bars, no seeds, no table of exact per-scenario results. The prose margins (+7.73% LoS, +9.90% beam) are not tied to the plotted bars in Figs. 3-4, and the percentages in Fig. 5 look like model confidence, not task accuracy—which is fine, but the paper should say so. I cannot tell if the gains are statistically meaningful.

3. The edge-resource claims (lower memory, energy, latency) are asserted, not measured. For a paper whose pitch is edge deployment, that is a missing piece. It also overclaims 'zero-shot'—it is really unseen-scenario generalization after fine-tuning on other cities, which is still decent but not zero-shot.

4. No comparison to other wireless-LLM approaches like RF-GPT, even though they cite it. The baselines are reasonable, but an LLM-based comparison would contextualize the contribution.

None of this is fatal. The core idea is plausible and the experiments, once cleaned up, could be solid. A careful referee should be able to fix the presentation and push for the missing ablations.

Who this is for: people working on foundation models for wireless and anyone building LLM-based physical-layer diagnostics. Send it to peer review, but expect heavy revision. I would not cite it yet—wait for the cleaned version.

Referee Report

3 major / 6 minor

Summary. The paper proposes PERA, a three-tier architecture that connects a pretrained Large Wireless AI Model (LWAM) sensory backbone to a pretrained LLM cognitive core through a trainable modality bridge. The architecture is motivated by a survey of physical-, network-, and application-layer AI techniques, and a two-stage multi-modal instruction-tuning protocol is introduced to align wireless feature representations with LLM token space. The empirical case study uses the DeepMIMO dataset for LoS/NLoS classification and beam-group prediction, reporting F1 gains over raw-CSI CNN and LWAM-token baselines across six unseen scenarios, and a single inference example is presented as evidence of human-readable, causally grounded rationales. The paper concludes that PERA replaces task-specific edge models while providing trust and explainability.

Significance. If fully supported, the paper would make a timely contribution to the emerging area of LLM-based wireless systems: the idea of feeding LWAM representations into a LoRA-adapted LLM via a lightweight projector, with auxiliary physical-label heads, is plausible and worth exploring. The two-stage training protocol and the evaluation on unseen DeepMIMO scenarios are concrete steps beyond pure position papers. However, the current evidence does not yet substantiate the central claims: the explainability/trust argument rests on a single hand-picked example, no faithfulness metric is defined, and the reported performance numbers are not clearly reconciled with the figures. The paper would be strengthened significantly by a rigorous ablation study, a quantitative faithfulness evaluation, and a complete per-scenario results table.

major comments (3)

[§IV-B, §IV-C, Fig. 5] The paper's stated differentiator is 'human-readable, causal explanations' (Abstract, §III-D, Conclusion). The only support is the single illustrative example in Fig. 5. Section IV-B states that auxiliary heads are included 'to preserve physical fidelity and prevent generative hallucinations,' but no ablation compares the 0.3:1.0:0.5 weighted objective against variants without the auxiliary heads or without the language loss, and no protocol is given for measuring whether a generated rationale is factually faithful to the channel state. This is load-bearing: the trust/explainability claim is a core contribution, not a peripheral feature. I recommend adding (i) an ablation over the auxiliary heads and loss weights; (ii) a quantitative faithfulness metric, e.g., agreement between rationale-cited token positions and actual L2-norm peaks/valleys over the test set, or a counterfactual test wh
[§IV-C, Fig. 3, Fig. 4] The claimed margins (+7.73%, +11.77%, +31.17%, +9.90%) are not traceable to the displayed data. The figures show only bars with approximate F1 values around 0.78–0.93, with no error bars, no per-scenario table, and no statement of whether the percentages are absolute or relative improvements. For example, a mean F1 of 0.93 versus 0.86 could be reported as '+7 percentage points' or '+8.1% relative', and Fig. 3 alone does not disambiguate this. Please provide exact per-scenario and mean±std values, define the percentage convention, and report significance tests. Without this, the central performance claim cannot be independently verified.
[§IV-A, §IV-B] The experimental protocol is underspecified in ways that affect fairness and reproducibility. It is unclear what 'MLP (LWM)' and 'Seq-CNN on LWM tokens' exactly are, whether they use the same auxiliary heads and training epochs, and how the 16 beam groups are defined. The five training and six inference scenarios are named, but key DeepMIMO parameters (scenario identifiers, numbers of users/BS antennas, SNR, sample counts) are missing. 'Zero-shot environmental adaptability' is used for held-out generalization; since the projector and LoRA adapters are trained on the five training scenarios, the LLM's contribution to adaptation is not isolated. Please clarify these points and, ideally, add an ablation that removes the LLM (keeping projector + auxiliary heads) to isolate the LLM's contribution to the reported F1 gains.

minor comments (6)

[Throughout] Notation is inconsistent: 'L WM' (Ref. [6]) versus 'L W AM' elsewhere; please unify and use a consistent non-breaking space.
[Fig. 3, Fig. 4, Fig. 2] Labels contain typos/artifacts: 'LoS/NLoS F1 by Scenari', 'Beam Predic(ion F1 by Scenario', 'MLP n L WM t k ens', and 'Casaule Rationale' in Fig. 2. Please correct.
[§IV-B] Hyperparameters (LoRA rank 16, alpha 32, MLP width 1024, loss weights 0.3:1.0:0.5) are given without any sensitivity analysis or selection rationale; one or two sentences explaining the choice would help.
[Fig. 5] The units and source of 'peak@122', 'valley@0', and 'CLS aggregate is 0.0523' are not defined. Please define the token indexing and the L2-norm aggregation so the example can be checked by readers.
[References] Some references appear very recent or hard to verify (e.g., [12] arXiv:2602.14833, [2] IEEE Wireless Communications 2026). Please confirm bibliographic details and provide DOIs where available.
[Reproducibility] No code, trained checkpoints, or dataset-generation scripts are provided. Given the novelty of the pipeline, a reproducibility statement or an anonymous code release would substantially increase confidence.

Circularity Check

0 steps flagged

No significant circularity: the empirical benchmark is external and the central performance claims do not reduce to the paper's own assumptions or fitted parameters.

full rationale

The paper's main derivation chain is the PERA architecture and its DeepMIMO case study. The claimed F1 gains (+7.73% LoS/NLoS over raw-CSI CNN, +9.90% beam over Seq-CNN on LWM tokens) are obtained by training on five scenarios and testing on six unseen scenarios; the test labels are external to the design, and no parameter is fitted to the test outcome and then relabeled as a prediction. The auxiliary heads and LoRA adapters are trained jointly on the task labels, which is ordinary supervised learning, not circularity. The 'trustworthy, human-readable rationales' claim is methodologically under-supported: Section IV-B asserts auxiliary pathways are included 'to preserve physical fidelity and prevent generative hallucinations,' and Fig. 5 provides a single illustrative rationale, but no faithfulness protocol, no grounding ablation, and no systematic evaluation of explanation correctness are provided. That is a soundness/evidence gap, not a circularity: the generated text is not equated with the supervision labels by construction, and the label predictions are measured independently. Self-citations [2], [5], [13] appear in the survey sections, but the load-bearing components cite external, independently checkable sources: LWM [6], Qwen-3, LLaVA MM-IT [14], and DeepMIMO [15]. No uniqueness theorem or prior self-result is used to force the architecture choice. The empirical benchmark is self-contained against external baselines, so no circular step can be exhibited under the required standard.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

No new physics or mathematical constants are invented; the free parameters are standard training hyperparameters whose values are chosen without ablation, and the axioms are domain assumptions about DeepMIMO validity, LWAM transferability, and LLM groundedness.

free parameters (3)

Loss weighting (lambda_LM, lambda_LoS, lambda_Beam) = 0.3:1.0:0.5 = 0.3, 1.0, 0.5
The relative weights for language modeling, LoS classification, and beam grouping appear hand-chosen with no ablation or sensitivity analysis (Section IV-B).
Curriculum: three stages of two epochs each, LR decaying 1e-3 to 1e-4 = 3x2 epochs
Training schedule is a hand-tuned hyperparameter choice; no sensitivity analysis or ablation is reported (Section IV-B).
LoRA rank 16, alpha 32 on QKVO projections = rank=16, alpha=32
Standard PEFT hyperparameters chosen without ablation; they affect the reported performance (Section IV-B).

axioms (3)

domain assumption DeepMIMO ray-tracing channels plus its LoS/NLoS and beam labels are a valid proxy for real 6G wireless environments.
The entire empirical validation (Section IV-A) relies on DeepMIMO simulations; generalization to real radio environments is asserted, not measured.
domain assumption The pretrained LWAM backbone (frozen) produces representations that preserve the physical information needed for LoS and beam prediction.
Layer 1 of the architecture is adopted from [6] (Section IV-A), itself an independent arXiv foundation model; PERA inherits its quality as an unverified input.
domain assumption A pretrained LLM (Qwen-3 Instruct), after LoRA tuning on paired channel-text samples, will ground its generated rationales in the wireless tokens rather than hallucinating.
The central explainability claim in Section IV-C assumes rationales are causally tied to channel features; no faithfulness/consistency metric is reported.

invented entities (2)

Network Pseudo-Tokens no independent evidence
purpose: Conceptual entity describing the projector output (Layer 2) that the LLM consumes as if it were text tokens.
This is a conceptual framing of a learned embedding sequence, not a new physical entity; its properties are not separately validated.
PERA Architecture (Perceive-Reason-Act) no independent evidence
purpose: The proposed system-level interface between wireless sensing, LLM reasoning, and control outputs.
It is the proposed architecture itself; its existence and performance are exactly what the paper claims to demonstrate.

pith-pipeline@v1.3.0-alltime-deepseek · 10939 in / 7287 out tokens · 60549 ms · 2026-08-01T19:34:48.685818+00:00 · methodology

0 comments

read the original abstract

The realization of next-generation (NG) networks hinges on a fundamental departure from preprogrammed protocol engineering towards a paradigm of self-consciously evolving, autonomous and trusted intelligence. While conventional machine learning (ML) has introduced localized automation, it remains inherently bounded by single-task processing pipelines incapable of handling complex cross-layer dynamics. As a partial remedy, large language models (LLMs) excel at generalized cognitive reasoning, but to a degree they remain detached from the rich modalities of wireless telemetry. As a solution, we unveil Generative Network Intelligence conceptualized via the Perceive-Reason-Act (PERA) paradigm. This paradigm treats the wireless channel and the underlying network states as a continuous, multimodal narrative. By synchronizing the perceptual grounding of Large Wireless AI Models (LWAMs) with the cognitive reasoning of LLMs, PERA heralds the era of native NG intelligence. Crucially, this unified intelligence replaces fragmented, task-specific edge models by an efficient multi-task architecture delivering the real-time control needed for supporting dynamic physical applications while reducing both the complexity and energy dissipation. Moreover, we contrast the structural limitations of traditional ML to generative paradigms, conceive agentic reasoning across a NG protocol stack, and detail a practical three-tier design specifically engineered for the resource-constrained wireless edge. This architectural paradigm serves as a foundational framework for realizing fully autonomous, embodied agentic AI in NG networks. To validate this vision, our case study evaluates link-state classification and beam prediction, demonstrating how grounding wireless telemetry within a cognitive engine delivers the transparent, human-readable rationales required for trusted physical-layer diagnostics and beam control.

Figures

Figures reproduced from arXiv: 2607.16896 by Lajos Hanzo, Melike Erol-Kantarci, Mohammad Farzanullah.

**Figure 1.** Figure 1: Framework for Embodied AI-driven 6G intelligence: (a) application domains and vertical deployments, and (b) synchronization of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: Overview of the PERA framework’s unification strategy, depicting the two-stage training protocol: Stage 1 (Feature Alignment) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison of F1 score for LoS/NLoS classification. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison of F1 score for beam prediction. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: Demonstration of the PERA architecture pipeline during [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 3 linked inside Pith

[1]

Wireless large AI model: Shaping the AI-native future of 6G and beyond,

F. Zhu, X. Wang, S. Jiang, X. Li, M. Zhang, Y. Chen, C. Huang, Z. Yang, X. Chen, Z. Zhang et al., “Wireless large AI model: Shaping the AI-native future of 6G and beyond,” arXiv preprint arXiv:2504.14653, 2025

Pith/arXiv arXiv 2025
[2]

Generative AI for Intent- Driven Network Management in 6G RAN: A Case Study on the Mamba Model,

M. A. Habib, M. Elsayed, Y. Ozcan, P. E. Iturria-Rivera, M. Bavand, and M. Erol-Kantarci, “Generative AI for Intent- Driven Network Management in 6G RAN: A Case Study on the Mamba Model,” IEEE Wireless Communications, pp. 1–8, 2026

2026
[3]

NetGPT: An AI-native network architecture for provisioning beyond personalized generative services,

Y. Chen, R. Li, Z. Zhao, C. Peng, J. Wu, E. Hossain, and H. Zhang, “NetGPT: An AI-native network architecture for provisioning beyond personalized generative services,” IEEE Network, vol. 38, no. 6, pp. 404–413, 2024. 9

2024
[4]

Large Language Models for Wireless Net- works: An Overview from the Prompt Engineering Perspective,

H. Zhou, C. Hu, D. Yuan, Y. Yuan, D. Wu, X. Chen, H. Tabas- sum, and X. Liu, “Large Language Models for Wireless Net- works: An Overview from the Prompt Engineering Perspective,” IEEE Wireless Communications, vol. 32, no. 4, pp. 98–106, 2025

2025
[5]

Wireless Multimodal Foundation Model (WMFM): Integrating Vision and Communication Modalities for 6G ISAC Systems,

M. Farzanullah, H. Zhang, A. B. Sediq, A. Afana, and M. Erol- Kantarci, “Wireless Multimodal Foundation Model (WMFM): Integrating Vision and Communication Modalities for 6G ISAC Systems,” arXiv preprint arXiv:2512.23897, 2025

arXiv 2025
[6]

Large wireless model (L WM): A foundation model for wireless channels,

S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (L WM): A foundation model for wireless channels,” arXiv preprint arXiv:2411.08872, 2024

Pith/arXiv arXiv 2024
[7]

IQFM—A Wireless Foundation Model for I/Q Streams in AI-Native 6G,

O. Mashaal and H. Abou-Zeid, “IQFM—A Wireless Foundation Model for I/Q Streams in AI-Native 6G,” IEEE Open Journal of the Communications Society, vol. 7, pp. 1426–1441, 2026

2026
[8]

Advanced end-to-end intent-driven dynamic network slicing: Challenges, solutions, and implementation,

L. Liu, A. Rajkumar, and Y. Jiang, “Advanced end-to-end intent-driven dynamic network slicing: Challenges, solutions, and implementation,” IEEE Communications Standards Mag- azine, pp. 1–8, 2025

2025
[9]

Generative diffusion models for wireless networks: Fundamen- tal, architecture, and state-of-the-art,

D. Fan, R. Meng, X. Xu, Y. Liu, G. Nan, C. Feng, S. Han, S. Gao, B. Xu, D. Niyato, T. Q. S. Quek, and P. Zhang, “Generative diffusion models for wireless networks: Fundamen- tal, architecture, and state-of-the-art,” IEEE Communications Surveys & Tutorials, vol. 28, pp. 5632–5677, 2026

2026
[10]

Edge large ai model empowered cognitive multimodal semantic communication system,

Y. Sun, S. Guo, X. Qiu, J. Chen, Y. Liu, R. Xin, and F. Qi, “Edge large ai model empowered cognitive multimodal semantic communication system,” in ICC 2025 - IEEE International Conference on Communications, 2025, pp. 3406–3411

2025
[11]

6g-bench: An open benchmark for semantic communication and network-level reasoning with foundation models in ai-native 6g networks,

M. A. Ferrag, A. Lakas, and M. Debbah, “6g-bench: An open benchmark for semantic communication and network-level reasoning with foundation models in ai-native 6g networks,” IEEE Open Journal of the Communications Society, vol. 7, pp. 3305–3330, 2026

2026
[12]

RF-GPT: Teaching AI to See the Wireless World,

H. Zou, Y. Tian, B. Wang, L. Bariah, S. Lasaulce, C. Huang, and M. Debbah, “RF-GPT: Teaching AI to See the Wireless World,” arXiv preprint arXiv:2602.14833, 2026

arXiv 2026
[13]

Chain-of-Thought Reasoning Enhances In- Context Learning for LLM-Based Mobile Traﬀic Prediction,

M. Ghadaksaz, M. Farzanullah, A. B. Sediq, A. Afana, and M. Erol-Kantarci, “Chain-of-Thought Reasoning Enhances In- Context Learning for LLM-Based Mobile Traﬀic Prediction,”
[14]

Visual instruction tuning,

H. Liu, C. Li, Q. Wu, and Y. J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, vol. 36, pp. 34 892–34 916, 2023

2023
[15]

DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,

A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,” in Proc. of Information Theory and Applications Workshop (ITA), San Diego, CA, Feb 2019, pp. 1–8

2019
[2026]

A vailable: https://arxiv.org/abs/2605.09260

[Online]. A vailable: https://arxiv.org/abs/2605.09260

Pith/arXiv arXiv

[1] [1]

Wireless large AI model: Shaping the AI-native future of 6G and beyond,

F. Zhu, X. Wang, S. Jiang, X. Li, M. Zhang, Y. Chen, C. Huang, Z. Yang, X. Chen, Z. Zhang et al., “Wireless large AI model: Shaping the AI-native future of 6G and beyond,” arXiv preprint arXiv:2504.14653, 2025

Pith/arXiv arXiv 2025

[2] [2]

Generative AI for Intent- Driven Network Management in 6G RAN: A Case Study on the Mamba Model,

M. A. Habib, M. Elsayed, Y. Ozcan, P. E. Iturria-Rivera, M. Bavand, and M. Erol-Kantarci, “Generative AI for Intent- Driven Network Management in 6G RAN: A Case Study on the Mamba Model,” IEEE Wireless Communications, pp. 1–8, 2026

2026

[3] [3]

NetGPT: An AI-native network architecture for provisioning beyond personalized generative services,

Y. Chen, R. Li, Z. Zhao, C. Peng, J. Wu, E. Hossain, and H. Zhang, “NetGPT: An AI-native network architecture for provisioning beyond personalized generative services,” IEEE Network, vol. 38, no. 6, pp. 404–413, 2024. 9

2024

[4] [4]

Large Language Models for Wireless Net- works: An Overview from the Prompt Engineering Perspective,

H. Zhou, C. Hu, D. Yuan, Y. Yuan, D. Wu, X. Chen, H. Tabas- sum, and X. Liu, “Large Language Models for Wireless Net- works: An Overview from the Prompt Engineering Perspective,” IEEE Wireless Communications, vol. 32, no. 4, pp. 98–106, 2025

2025

[5] [5]

Wireless Multimodal Foundation Model (WMFM): Integrating Vision and Communication Modalities for 6G ISAC Systems,

M. Farzanullah, H. Zhang, A. B. Sediq, A. Afana, and M. Erol- Kantarci, “Wireless Multimodal Foundation Model (WMFM): Integrating Vision and Communication Modalities for 6G ISAC Systems,” arXiv preprint arXiv:2512.23897, 2025

arXiv 2025

[6] [6]

Large wireless model (L WM): A foundation model for wireless channels,

S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (L WM): A foundation model for wireless channels,” arXiv preprint arXiv:2411.08872, 2024

Pith/arXiv arXiv 2024

[7] [7]

IQFM—A Wireless Foundation Model for I/Q Streams in AI-Native 6G,

O. Mashaal and H. Abou-Zeid, “IQFM—A Wireless Foundation Model for I/Q Streams in AI-Native 6G,” IEEE Open Journal of the Communications Society, vol. 7, pp. 1426–1441, 2026

2026

[8] [8]

Advanced end-to-end intent-driven dynamic network slicing: Challenges, solutions, and implementation,

L. Liu, A. Rajkumar, and Y. Jiang, “Advanced end-to-end intent-driven dynamic network slicing: Challenges, solutions, and implementation,” IEEE Communications Standards Mag- azine, pp. 1–8, 2025

2025

[9] [9]

Generative diffusion models for wireless networks: Fundamen- tal, architecture, and state-of-the-art,

D. Fan, R. Meng, X. Xu, Y. Liu, G. Nan, C. Feng, S. Han, S. Gao, B. Xu, D. Niyato, T. Q. S. Quek, and P. Zhang, “Generative diffusion models for wireless networks: Fundamen- tal, architecture, and state-of-the-art,” IEEE Communications Surveys & Tutorials, vol. 28, pp. 5632–5677, 2026

2026

[10] [10]

Edge large ai model empowered cognitive multimodal semantic communication system,

Y. Sun, S. Guo, X. Qiu, J. Chen, Y. Liu, R. Xin, and F. Qi, “Edge large ai model empowered cognitive multimodal semantic communication system,” in ICC 2025 - IEEE International Conference on Communications, 2025, pp. 3406–3411

2025

[11] [11]

6g-bench: An open benchmark for semantic communication and network-level reasoning with foundation models in ai-native 6g networks,

M. A. Ferrag, A. Lakas, and M. Debbah, “6g-bench: An open benchmark for semantic communication and network-level reasoning with foundation models in ai-native 6g networks,” IEEE Open Journal of the Communications Society, vol. 7, pp. 3305–3330, 2026

2026

[12] [12]

RF-GPT: Teaching AI to See the Wireless World,

H. Zou, Y. Tian, B. Wang, L. Bariah, S. Lasaulce, C. Huang, and M. Debbah, “RF-GPT: Teaching AI to See the Wireless World,” arXiv preprint arXiv:2602.14833, 2026

arXiv 2026

[13] [13]

Chain-of-Thought Reasoning Enhances In- Context Learning for LLM-Based Mobile Traﬀic Prediction,

M. Ghadaksaz, M. Farzanullah, A. B. Sediq, A. Afana, and M. Erol-Kantarci, “Chain-of-Thought Reasoning Enhances In- Context Learning for LLM-Based Mobile Traﬀic Prediction,”

[14] [14]

Visual instruction tuning,

H. Liu, C. Li, Q. Wu, and Y. J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, vol. 36, pp. 34 892–34 916, 2023

2023

[15] [15]

DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,

A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications,” in Proc. of Information Theory and Applications Workshop (ITA), San Diego, CA, Feb 2019, pp. 1–8

2019

[16] [2026]

A vailable: https://arxiv.org/abs/2605.09260

[Online]. A vailable: https://arxiv.org/abs/2605.09260

Pith/arXiv arXiv