ORAN-GUIDE: RAG-Driven Prompt Learning for LLM-Augmented Reinforcement Learning in O-RAN Network Slicing

Fatemeh Afghah; Fatemeh Lotfi; Hossein Rajoli

REVIEW 4 major objections 6 minor 1 cited by

ORAN-GUIDE: RAG-Driven Prompt Learning for LLM-Augmented Reinforcement Learning in O-RAN Network Slicing

T0 review · 4 major / 6 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 A dual-LLM prompt-fusion state representation improves sample efficiency, convergence, and generalization in multi-agent reinforcement learning for O-RAN network slicing.

desk verdict Sensible modular architecture, but the paper never shows the ORANSight prompt adds information beyond the raw numeric state, so the central claim is unsupported. read the letter →

arxiv 2506.00576 v1 pith:3RNY5WQ2 submitted 2025-05-31 cs.LG cs.AI

Fatemeh Lotfi , Hossein Rajoli , Fatemeh Afghah This is my paper

classification cs.LGcs.AI

keywords O-RANnetworkslicingmulti-agentreinforcementlearninglargelanguagemodelsretrieval-augmentedgenerationprompttuningsoftactor-criticknowledgedistillationstaterepresentation

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 tries to establish that multi-agent reinforcement learning for network slicing in Open RAN can be made more sample-efficient and better at generalizing by giving each agent a semantic view of the radio state, produced by two language models working together. The proposed pipeline, ORAN-GUIDE, sends current QoS levels, resource-block allocations, and user counts through a domain-specific language model (ORANSight) that turns them into a textual prompt; those words are concatenated with trainable prompt tokens and encoded by a frozen GPT-2 model into a state embedding for the RL policy. The paper argues that this dual-prompt fusion injects domain knowledge while keeping the large models frozen, so the system stays light enough for edge deployment. In simulations, it reports faster convergence, higher cumulative reward, lower run-to-run variability, and better per-slice QoS than standard MARL or single-LLM baselines. A sympathetic reader would care because the design offers a modular way to add domain expertise to RL without fine-tuning large models.

What carries the argument

The central mechanism is the State Representation Module (SRM), a prompt-fusion encoder that carries the entire argument. It generates a domain prompt $P_{\text{domain}}$ from ORANSight based on the live telemetry, concatenates it with a trainable prompt $P_{\text{learnable}}$, passes the token sequence through a frozen GPT-2 encoder to obtain a semantic embedding, and then uses two lightweight adapter networks, $F_{c1}$ and $F_{c2}$, to align this embedding with the raw numeric state for the SAC actor-critic agents. A knowledge-distillation loss aligns GPT-2's representation with ORANSight's, letting the learnable tokens absorb domain knowledge while the large models stay frozen.

What would settle it

Train ORAN-GUIDE and a control variant in which ORANSight is replaced by a fixed template that prints the same numeric state in the same sentence structure, while keeping the learnable tokens, adapters, and SAC loop unchanged; if the control's cumulative reward and RPI stay within noise of ORAN-GUIDE, the reported gains come from the learnable prompt pipeline rather than from the domain LLM's semantic content.

Watch

Extended reading notes

Core claim

The core claim is that replacing raw numerical state vectors with representations produced by a dual-LLM prompt-fusion module improves MARL for O-RAN slicing. At each step, ORANSight—a language model fine-tuned on O-RAN control and configuration data—reads the current network state and generates a context-aware prompt describing QoS levels, allocated resource blocks, and connected users. A frozen GPT-2 encoder receives this prompt concatenated with learnable tokens optimized by the RL loss, and outputs a semantic embedding; two adapter networks map this embedding and the raw state into a shared latent space consumed by distributed Soft Actor-Critic agents with a centralized critic. The paper reports that ORAN-GUIDE achieves the highest median cumulative reward with lower variability across runs, the best Relative Performance Improvement (14.74% over the MARL baseline), faster convergence, and higher per-user throughput across all three slices in the ablation study.

Load-bearing premise

The framework assumes that ORANSight's textual prompt conveys information to the RL agent beyond the same values already present in the raw numeric state.

Editorial extensions

If this is right

ORAN-GUIDE is claimed to converge faster and reach higher average cumulative rewards than plain MARL, ORANSight-only MARL, and GPT-only prompt-augmented MARL.
The dual-LLM design separates domain-knowledge generation from policy execution, so the domain model can be swapped without retraining the whole decision pipeline.
Knowledge distillation lets the learnable prompts absorb ORANSight's domain knowledge, so the lightweight GPT module can run at the edge without centralized ORANSight inference.
The number of learnable tokens is a tunable hyperparameter: too few underfit and too many overfit, with an optimal range reported in the experiments.
Per-user throughput improves across eMBB, mMTC, and URLLC slices, with ORAN-GUIDE giving the highest throughput distribution among the compared methods.

Reading between the lines

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

A control experiment replacing ORANSight's prompt with a plain templated restatement of the same numbers would isolate whether the domain LLM or the learnable prompt pipeline is the source of the gains; the paper does not report this control.
The same SRM recipe could transfer to other partially observable control problems with domain documentation, such as power-grid management or autonomous driving, whenever a domain-specific LLM is available.
A capacity-matched baseline, plain MARL with a larger policy network, would test whether the gains exceed what extra representational capacity alone would buy; the paper does not include this control.

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

The modular design—separating a domain-knowledge generator from a frozen policy encoder via learnable tokens and distillation—is a genuinely reasonable idea worth pursuing. But this manuscript doesn't give you a reason to believe the ORANSight prompt is doing the work the paper claims.

What's new: the specific combination of ORANSight as runtime prompt generator, GPT-2 as frozen semantic encoder, learnable tokens, and distillation. Each piece is known, but the packaging is theirs. The system model and problem formulation are clear, and the ablation against single-LLM variants is an appropriate way to isolate the contribution.

The problems are in the evidence. The only prompt template shown (Fig. 2) is a direct f-string serialization of QoS values, RB counts, and user counts. That doesn't support "context-rich" or "semantically enriched" prompts. The paper calls the pipeline RAG but never describes a retrieval step, a knowledge base, an embedding index, or how the offline dataset and cosine-similarity block in Fig. 2 are used. The distillation loss Ldistill in Algorithm 1 is a KL divergence between a GPT-2 sequence embedding and an ORANSight output with no specified alignment. Hyperparameters—reward coefficients, priority weights, token count, distillation λ—are largely unreported, and the token count is explicitly tuned on the evaluation metric in Fig. 6. That selection-on-eval issue is a real threat.

The deeper confound: ORAN-GUIDE has two LLMs plus learnable tokens and adapters, while baselines use one LLM or no prompt alignment. The RPI gains (14.74% vs 9.0%, 2.16%, 0.84%) could simply reflect additional capacity and longer token sequences. The ablation doesn't control for that, so the central attribution—domain knowledge injection—is untested. There are also no error bars on Table III or significance tests; Fig. 4 shows distributions but they're not tied to the headline metrics.

On the positive side, the paper is honest about building on the authors' prior ICC work [14] and ORANSight [20], and the writing is clear. No code or data is provided, so reproducibility is near zero.

Who gets value? A researcher working on LLM-augmented RL for wireless who wants a catalogue of pitfalls in prompt pipelines. Not someone looking for a validated method. A serious referee could help the authors fix the missing details, but as it stands the central claims aren't supported by the implementation shown.

Recommendation: if this crosses your desk, send it to review only if you have time for a demanding round; the architecture is worth a look but the paper needs major revision. My own verdict: reject as-is.

Referee Report

4 major / 6 minor

Summary. The paper proposes ORAN-GUIDE, a dual-LLM framework for multi-agent reinforcement learning (MARL) in O-RAN network slicing. A domain-specific LLM (ORANSight) is said to perform retrieval-augmented generation (RAG) to produce context-rich prompts from the current network state; these are fused with learnable prompt tokens and encoded by a frozen GPT-2 model, whose outputs are used as semantic state representations for Soft Actor-Critic agents. The authors evaluate the method in a simulated O-RAN environment with three slices (eMBB, mMTC, URLLC), reporting improved convergence, sample efficiency (via a Relative Performance Improvement metric), and per-slice QoS compared to several LLM-augmented and plain MARL baselines.

Significance. If the results are valid, the proposed separation of domain-knowledge generation (ORANSight) from task-level encoding (GPT-2 + learnable prompts) would be a useful architectural idea for injecting domain knowledge into MARL for wireless control. The paper gives a clear system model and a well-defined MDP, and the ablation family is reasonable in principle. However, the central claim of 'semantically enriched state representations' is not supported by the implementation details: the only prompt template shown is a direct serialization of numeric state variables, and the RAG mechanism is never specified. The empirical comparisons also do not control for model capacity or statistical variability. The paper does not provide reproducible code or machine-checked proofs, so the contribution currently rests on an under-specified and confounded evaluation.

major comments (4)

[Section IV-D and Fig. 2] The paper's central claim is that ORANSight generates 'context-rich' prompts that add semantic value beyond the raw state. However, the only prompt template shown in Fig. 2 is a direct f-string serialization of state_action_list values (QoS numbers, RB counts, user counts) with no retrieved documents, no domain inference, and no additional guidance. Section IV-D states that ORANSight 'performs dynamic RAG,' but the retrieval step, knowledge base, embedding index, and retrieved-context input are never described; the 'offline dataset' and 'cosine similarity' blocks in Fig. 2 are not explained anywhere in the text. As written, the manuscript does not establish that Pdomain carries any information beyond the raw numeric state, which is load-bearing for the claimed semantic enrichment.
[Table III and Section V.D] The ablation study does not control for model capacity or tunable parameters. ORAN-GUIDE uses two LLMs (ORANSight and GPT-2) plus learnable tokens and adapter networks, whereas the baselines use only one LLM or omit prompt alignment. Section V.D shows that the number of learnable tokens is tuned on the evaluation metric ('the token count should be treated as a tunable hyperparameter'), introducing potential selection bias. The reported RPI gains (14.74% vs 9.0%, 2.16%, 0.84%) can therefore be explained by additional capacity, longer token sequences, or hyperparameter tuning, rather than by the proposed RAG-driven semantic prompts. No error bars or significance tests are provided for Table III or Figs. 5-6.
[Algorithm 1, line 14] The distillation loss Ldistill = KL(Mg(Xt)∥Md(st)) is undefined. Mg(Xt) is the token-sequence embedding produced by GPT-2, while Md(st) is described as ORANSight's state-derived prompt; the paper does not specify how these two objects are converted into probability distributions over a common space, nor how a KL divergence between them is computed. This makes the knowledge-distillation mechanism untestable and the 'distillation coefficient λ' in the combined loss (line 19) meaningless as specified. Additionally, Section IV.F says the adapters Fc1 and Fc2 are 'trained offline using paired data,' but Algorithm 1 includes the distillation loss in the online RL loop; the relationship between these two training stages is not clarified.
[Section V.B and Fig. 4] The claim that ORAN-GUIDE 'exhibits lower variability compared to other methods' is not supported by the reported statistical evidence. The number of random seeds or independent runs used to produce Fig. 4 is not stated, Fig. 5 appears to show a single training run without confidence bands, and no significance tests are reported. Without this information, the 'robustness' claim and the comparison of median performance are not verifiable.

minor comments (6)

[Section III.D and Section IV.B] The symbol γ is used both as the reward-shaping coefficient in Eq. (7) and as the discount factor in the MDP definition; this notational collision should be fixed by renaming one of them.
[Eq. (6d)] The constraint bl,k ≤ 1 + λl max(0, Σl bl,k − 1) appears to be an identity that does not impose a meaningful RB-sharing limit; please clarify the intended relaxation.
[Fig. 2] The 'offline dataset' and 'cosine similarity' blocks in Fig. 2 are never described in the body of the paper; either explain these components or remove them from the architecture diagram.
[References] Reference [29] appears to duplicate [14] with a truncated title ('LLM-augmented deep reinforcement learning for dynamic o-ran network slicing, environments, vol. 2, p. 4'); please correct the reference list.
[Section V.A] The experimental setup does not state how many random seeds or independent runs were used for any of the reported results; please add this information to allow reproducibility.
[Section V.A] There is a typo in the description of the proposed method: 'integerating' should be 'integrating'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are empirical simulator comparisons and do not reduce by construction to the method's inputs.

full rationale

ORAN-GUIDE's central claim is an empirical performance improvement measured in a simulated O-RAN environment. The accumulated-reward comparison, RPI metric, and ablation results are computed from environment interaction, not derived from the method's own definitions or fitted parameters. The paper does not define the target result in terms of its inputs: the RPI formula (Eq. 12) is a generic relative-improvement metric applied to measured rewards, and there is no equation in which ORAN-GUIDE's advantage is forced by construction. Self-citations to the authors' prior work [14] appear for the adapter networks, baseline variants, and user-mobility model, but these are reused components rather than load-bearing proof of the superiority claim. The domain LLM ORANSight is cited from external work [20], and GPT-2 is an external model [21], so the architecture's novelty does not rest on an unverified self-citation chain. The paper's implemented prompt template is only a textual serialization of the numeric state, and the described RAG retrieval step is not concretely specified; these are substantive weaknesses regarding whether the claimed semantic enrichment is actually tested, but they are not circularity. Similarly, the undefined distillation loss and the hyperparameter tuning of learnable tokens are implementation and confound concerns, not reductions of output to input. No step in the derivation chain equates a prediction to a fitted quantity or imports a uniqueness result from the authors' prior work. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical or conceptual entities are postulated. The learnable prompt tokens are trainable parameters, not separate entities, and ORANSight and GPT-2 are externally pretrained models. The 'RAG' pipeline is claimed but not concretely instantiated in the text.

free parameters (5)

reward shaping coefficients (beta, gamma, alpha, delta, margin, thr) = not reported
Eqs. (7)-(9) define the RL reward with at least six coefficients; all reported QoS gains are maximized with respect to this hand-shaped reward, and the values are never given.
slice priority weights w_l = not reported
Eq. (6a) weights each slice's utility; the values are not stated, yet they directly affect the optimal allocation and thus the comparison.
learnable token count = optimal value selected in Fig. 6 (marked red)
Section V.D treats the number of learnable tokens as a tunable hyperparameter and selects the best performing one, so the reported ORAN-GUIDE results include this tuning.
distillation coefficient lambda = not reported
Algorithm 1 combines L_RL and lambda * L_distill, but lambda is never assigned, so the strength of the distillation loss is unspecified.
RB sharing relaxation lambda_l = not reported
Constraint (6d) includes a relaxation term lambda_l, which is never assigned and could affect feasibility and the learned allocations.

assumptions (6)

domain assumption Eq. (1) Shannon-type capacity with AWGN and inter-cell interference models the OFDM physical layer.
Section III.B states the rate formula as given; it is a standard approximation, not derived in the paper, and the simulator is built on it.
ad hoc to paper The piecewise reward in Eqs. (7)-(9) faithfully encodes the three slice QoS objectives.
The sigmoid/exponential reward is introduced without justification or parameter values; all reported improvements are measured under this reward.
domain assumption ORANSight-2.0's pretrained O-RAN knowledge creates prompts that are useful for RL beyond the raw state.
The framework's central mechanism assumes the domain model adds semantic value, but the only prompt template shown is a serialization of state values (Fig. 2).
domain assumption A frozen GPT-2, augmented with learnable prompt tokens and two adapters, yields a good state representation for RL.
The paper assumes the frozen LLM's hidden states transfer to the RL task; no analysis or comparison of alternative encoders is provided.
standard math Soft Actor-Critic (Eqs. 10-11) provides a convergent policy optimization procedure.
SAC is used as an off-the-shelf algorithm without proof; this is standard background and not a source of novelty.
domain assumption The simulated environment (mobility model from [14], 3GPP PHY from [33]) represents realistic O-RAN dynamics.
The simulator is assembled from the authors' prior work and a 3GPP document; its validity as a proxy for real O-RAN is not tested.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ORAN-GUIDE: RAG-Driven Prompt Learning for LLM-Augmented Reinforcement Learning in O-RAN Network Slicing." pith.science (2026). https://pith.science/paper/3RNY5WQ2

@misc{pith2026250600576,
  author       = {Pith},
  title        = {Pith review of: ORAN-GUIDE: RAG-Driven Prompt Learning for LLM-Augmented Reinforcement Learning in O-RAN Network Slicing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3RNY5WQ2}},
  note         = {Machine review of arXiv:2506.00576}
}

read the original abstract

Advanced wireless networks must support highly dynamic and heterogeneous service demands. Open Radio Access Network (O-RAN) architecture enables this flexibility by adopting modular, disaggregated components, such as the RAN Intelligent Controller (RIC), Centralized Unit (CU), and Distributed Unit (DU), that can support intelligent control via machine learning (ML). While deep reinforcement learning (DRL) is a powerful tool for managing dynamic resource allocation and slicing, it often struggles to process raw, unstructured input like RF features, QoS metrics, and traffic trends. These limitations hinder policy generalization and decision efficiency in partially observable and evolving environments. To address this, we propose \textit{ORAN-GUIDE}, a dual-LLM framework that enhances multi-agent RL (MARL) with task-relevant, semantically enriched state representations. The architecture employs a domain-specific language model, ORANSight, pretrained on O-RAN control and configuration data, to generate structured, context-aware prompts. These prompts are fused with learnable tokens and passed to a frozen GPT-based encoder that outputs high-level semantic representations for DRL agents. This design adopts a retrieval-augmented generation (RAG) style pipeline tailored for technical decision-making in wireless systems. Experimental results show that ORAN-GUIDE improves sample efficiency, policy convergence, and performance generalization over standard MARL and single-LLM baselines.

Figures

Figures reproduced from arXiv: 2506.00576 by the authors.

**Figure 1.** Overview of the ORAN-GUIDE system architecture showing distributed DRL agents coordinated by a centralized critic, with promptaugmented State Representation Modules (SRMs). cognition in AI: a) ORANSight serves as a “planner” or domain expert, b) GPT acts as the “controller” or task executor. The result is a RAG-inspired decision-making framework, tailored for DRL in technical wireless environments. It leverages a … view at source ↗

**Figure 2.** SRM design illustrating how domain-specific and learnable prompts are fused and encoded by GPT to produce semantic state embeddings. This hybrid setup improves policy stability, responsiveness, and interpretability in real-time O-RAN control. IV. THE ORAN-GUIDE FRAMEWORK A. Overview of the Proposed Architecture In this framework, we leverage a domain-specific large language model, ORANSight, to implement a RAG-styl… view at source ↗

**Figure 3.** Training workflow of ORAN-GUIDE, integrating dual LLM prompts into DRL via a soft actor-critic loop, with distillation and prompt fusion for adaptive policy learning. trends, and interference conditions. They provide scenarioaware guidance that enhances decision flexibility and contextual awareness. 2) Learnable Prompts (Plearnable): These are trainable prompt tokens that embed task-specific knowledge and evolve d… view at source ↗

Figures from the paper (3 more)

**Figure 6.** Figure 6: Comparison of maximum reward gains under varying numbers of learnable context. TABLE III: Ablation Study Results Model Version QoS (eMBB) QoS (mMTC) QoS (URLLC) RPI ORAN-GUIDE 87.7% 62.0% 66.34% 14.74% ORANSight PA-MARL 85.3% 46.84% 25.42% 9.0% GPT PA-MARL 34.4% 21.57…

**Figure 5.** Figure 5: Comparison of average cumulative rewards across methods. specific LLM with prompt-aligned SRMs to provide contextaware insights at both the actor and critic levels; (3) GPT prompt augmented MARL (GPT PA-MARL), which replaces ORANSight with GPT-2 while retaining prompt…

**Figure 7.** Figure 7: CDF of UEs throughput across the network. F. Network users’ QoE [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

AI Reasoning for Wireless Communications and Networking: A Survey and Perspectives
cs.NI 2025-09 conditional novelty 4.0 of 10

A survey that organizes LLM and AI reasoning methods into a taxonomy and maps them onto the physical, link, network, transport, and application layers of wireless networks.

Reference graph

Works this paper leans on

33 extracted references · 18 canonical work pages · cited by 1 Pith paper

[14]

Llm-augmented deep reinforcement learning for dynamic o-ran network slicing,

F. Lotfi, H. Rajoli, and F. Afghah, “Llm-augmented deep reinforcement learning for dynamic o-ran network slicing,” ICC 2025-2025 IEEE International Conference on Communications (ICC) , 2025

work page 2025
[20]

Oransight-2.0: Foundational llms for o-ran,

P. Gajjar and V . K. Shah, “Oransight-2.0: Foundational llms for o-ran,” arXiv preprint arXiv:2503.05200 , 2025

arXiv 2025
[1]

Near-real-time ran intelligent controller use cases and requirements,

O-RAN WG3, “Near-real-time ran intelligent controller use cases and requirements,” O-RAN Alliance, Tech. Rep. O-RAN.WG3.UCR-R003- v6.00 TS, 2024

work page 2024
[2]

A Comprehensive Tutorial and Survey of O-RAN: Exploring Slicing-aware Architecture, Deployment Options, Use Cases, and Challenges

K. Alam et al. , “A comprehensive overview and survey of o-ran: Ex- ploring slicing-aware architecture, deployment options, and use cases,” arXiv preprint arXiv:2405.03555 , 2024

work page Pith review arXiv 2024
[3]

Study on enhanced access to and support of network slices,

G. T. . version 18.2.0 Release 18, “Study on enhanced access to and support of network slices,” Tech. Rep, no. 5, 2021

work page 2021
[4]

Un- derstanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,

M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Un- derstanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,” arXiv preprint arXiv:2202.01032 , 2022

arXiv 2022
[5]

Evolutionary deep reinforcement learning for dynamic slice management in O-RAN,

F. Lotfi, O. Semiari, and F. Afghah, “Evolutionary deep reinforcement learning for dynamic slice management in O-RAN,” in 2022 IEEE Globecom Workshops (GC Wkshps) . IEEE, 2022, pp. 227–232

work page 2022
[6]

Open RAN LSTM traffic prediction and slice management using deep reinforcement learning,

F. Lotfi and F. Afghah, “Open RAN LSTM traffic prediction and slice management using deep reinforcement learning,” in 2023 57th Asilomar Conference on Signals, Systems, and Computers , 2023, pp. 646–650

work page 2023

Show all 33 references

[7]

Attention-based open ran slice management using deep reinforcement learning,

F. Lotfi, F. Afghah, and J. Ashdown, “Attention-based open ran slice management using deep reinforcement learning,” in GLOBECOM 2023 - 2023 IEEE Global Communications Conference, 2023, pp. 6328–6333

2023
[8]

Meta reinforcement learning approach for adaptive resource optimization in o-ran,

F. Lotfi and F. Afghah, “Meta reinforcement learning approach for adaptive resource optimization in o-ran,” 2024. [Online]. Available: https://arxiv.org/abs/2410.03737

2024 arXiv
[9]

Federated deep reinforce- ment learning for resource allocation in O-RAN slicing,

H. Zhang, H. Zhou, and M. Erol-Kantarci, “Federated deep reinforce- ment learning for resource allocation in O-RAN slicing,” in GLOBE- COM IEEE Global Communications Conference . IEEE, 2022, pp. 958–963

2022
[10]

Resource management in wireless networks via multi-agent deep reinforcement learning,

N. Naderializadeh et al. , “Resource management in wireless networks via multi-agent deep reinforcement learning,” IEEE Transactions on Wireless Communications, vol. 20, no. 6, pp. 3507–3523, 2021

2021
[11]

Large generative ai models for telecom: The next big thing?

L. Bariah, Q. Zhao, H. Zou, Y . Tian, F. Bader, and M. Debbah, “Large generative ai models for telecom: The next big thing?” IEEE Communications Magazine, 2024

2024
[12]

Understanding telecom language through large language models,

L. Bariah, H. Zou, Q. Zhao, B. Mouhouche, F. Bader, and M. Debbah, “Understanding telecom language through large language models,” in GLOBECOM 2023-2023 IEEE Global Communications Conference . IEEE, 2023, pp. 6542–6547

2023
[13]

Communication and control co-design in 6g: Sequential decision-making with llms,

X. Chen et al., “Communication and control co-design in 6g: Sequential decision-making with llms,” arXiv preprint arXiv:2407.06227 , 2024

2024 arXiv
[15]

Llm-based intent processing and network optimiza- tion using attention-based hierarchical reinforcement learning,

M. Habib et al. , “Llm-based intent processing and network optimiza- tion using attention-based hierarchical reinforcement learning,” arXiv preprint arXiv:2406.06059, 2024

2024 arXiv
[16]

Controlling large language model-based agents for large-scale decision-making: An actor-critic approach,

B. Zhang et al. , “Controlling large language model-based agents for large-scale decision-making: An actor-critic approach,” arXiv preprint arXiv:2311.13884, 2023

2023 arXiv
[17]

Llm-empowered state representation for reinforcement learning,

B. Wang, Y . Qu, Y . Jiang, J. Shao, C. Liu, W. Yang, and X. Ji, “Llm-empowered state representation for reinforcement learning,” arXiv preprint arXiv:2407.13237, 2024

2024 arXiv
[18]

Lora: Low-rank adaptation of large language models

E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

2022
[19]

Parameter-efficient transfer learning for nlp,

N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-efficient transfer learning for nlp,” in International conference on machine learning . PMLR, 2019, pp. 2790–2799

2019
[21]

Language models are unsupervised multitask learners,

A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog , vol. 1, no. 8, p. 9, 2019

2019
[22]

Interactive ai with retrieval-augmented generation for next generation networking,

R. Zhang, H. Du, Y . Liu, D. Niyato, J. Kang, S. Sun, X. Shen, and H. V . Poor, “Interactive ai with retrieval-augmented generation for next generation networking,” IEEE Network , vol. 38, no. 6, pp. 414–424, 2024

2024
[23]

Toward effective retrieval augmented generative services in 6g networks,

X. Huang, Y . Tang, J. Li, N. Zhang, and X. S. Shen, “Toward effective retrieval augmented generative services in 6g networks,” IEEE Network, 2024

2024
[24]

Retrieval-augmented generation for mobile edge computing via large language model,

R. Ren, Y . Wu, X. Zhang, J. Ren, Y . Shen, S. Wang, and K.-F. Tsang, “Retrieval-augmented generation for mobile edge computing via large language model,” arXiv preprint arXiv:2412.20820 , 2024

2024 arXiv
[25]

Enwar: A rag-empowered multi-modal llm framework for wireless environment perception,

A. M. Nazar, A. Celik, M. Y . Selim, A. Abdallah, D. Qiao, and A. M. Eltawil, “Enwar: A rag-empowered multi-modal llm framework for wireless environment perception,” arXiv preprint arXiv:2410.18104, 2024

2024 arXiv
[26]

An autonomous network orchestration framework integrating large language models with continual reinforce- ment learning,

M. Shokrnezhad and T. Taleb, “An autonomous network orchestration framework integrating large language models with continual reinforce- ment learning,” arXiv preprint arXiv:2502.16198 , 2025

2025 arXiv
[27]

Wirelessllm: Empowering large language models towards wireless intelligence,

J. Shao, J. Tong, Q. Wu, W. Guo, Z. Li, Z. Lin, and J. Zhang, “Wirelessllm: Empowering large language models towards wireless intelligence,” arXiv preprint arXiv:2405.17053 , 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

2024 arXiv
[28]

Telecomrag: Taming telecom standards with retrieval augmented generation and llms,

G. M. Yilma, J. A. Ayala-Romero, A. Garcia-Saavedra, and X. Costa- Perez, “Telecomrag: Taming telecom standards with retrieval augmented generation and llms,” ACM SIGCOMM Computer Communication Re- view, vol. 54, no. 3, pp. 18–23, 2025

2025
[29]

Llm-augmented deep reinforcement learning for dynamic o-ran network slicing,

F. Lotfi, H. Rajoli, and F. Afghah, “Llm-augmented deep reinforcement learning for dynamic o-ran network slicing,” environments, vol. 2, p. 4
[30]

Oran-bench-13k: An open source benchmark for assessing llms in open radio access networks,

P. Gajjar and V . K. Shah, “Oran-bench-13k: An open source benchmark for assessing llms in open radio access networks,” arXiv preprint arXiv:2407.06245, 2024

2024 arXiv
[31]

The power of scale for parameter-efficient prompt tuning,

B. Lester, R. Al-Rfou, and N. Constant, “The power of scale for parameter-efficient prompt tuning,” arXiv preprint arXiv:2104.08691 , 2021

2021 arXiv
[32]

Prefix-tuning: Optimizing continuous prompts for generation,

X. L. Li and P. Liang, “Prefix-tuning: Optimizing continuous prompts for generation,” arXiv preprint arXiv:2101.00190 , 2021

2021 arXiv
[33]

5G; NR; physical channels and modulation,

G. T. . version 15.1.0 Release 15, “5G; NR; physical channels and modulation,” Tech. Spec., Jul. 2018

2018

Pith tools

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

[1] [14]

Llm-augmented deep reinforcement learning for dynamic o-ran network slicing,

F. Lotfi, H. Rajoli, and F. Afghah, “Llm-augmented deep reinforcement learning for dynamic o-ran network slicing,” ICC 2025-2025 IEEE International Conference on Communications (ICC) , 2025

work page 2025

[2] [20]

Oransight-2.0: Foundational llms for o-ran,

P. Gajjar and V . K. Shah, “Oransight-2.0: Foundational llms for o-ran,” arXiv preprint arXiv:2503.05200 , 2025

arXiv 2025

[3] [1]

Near-real-time ran intelligent controller use cases and requirements,

O-RAN WG3, “Near-real-time ran intelligent controller use cases and requirements,” O-RAN Alliance, Tech. Rep. O-RAN.WG3.UCR-R003- v6.00 TS, 2024

work page 2024

[4] [2]

A Comprehensive Tutorial and Survey of O-RAN: Exploring Slicing-aware Architecture, Deployment Options, Use Cases, and Challenges

K. Alam et al. , “A comprehensive overview and survey of o-ran: Ex- ploring slicing-aware architecture, deployment options, and use cases,” arXiv preprint arXiv:2405.03555 , 2024

work page Pith review arXiv 2024

[5] [3]

Study on enhanced access to and support of network slices,

G. T. . version 18.2.0 Release 18, “Study on enhanced access to and support of network slices,” Tech. Rep, no. 5, 2021

work page 2021

[6] [4]

Un- derstanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,

M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Un- derstanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,” arXiv preprint arXiv:2202.01032 , 2022

arXiv 2022

[7] [5]

Evolutionary deep reinforcement learning for dynamic slice management in O-RAN,

F. Lotfi, O. Semiari, and F. Afghah, “Evolutionary deep reinforcement learning for dynamic slice management in O-RAN,” in 2022 IEEE Globecom Workshops (GC Wkshps) . IEEE, 2022, pp. 227–232

work page 2022

[8] [6]

Open RAN LSTM traffic prediction and slice management using deep reinforcement learning,

F. Lotfi and F. Afghah, “Open RAN LSTM traffic prediction and slice management using deep reinforcement learning,” in 2023 57th Asilomar Conference on Signals, Systems, and Computers , 2023, pp. 646–650

work page 2023