AI2MMUM: AI-AI Oriented Multi-Modal Universal Model Leveraging Telecom Domain Large Model

Chenhui Ye; Dazhi He; Fangkun Liu; Jiang Chang; Liyu Cai; Tianyu Jiao; Wenjun Zhang; Yihang Huang; Yijia Feng; Yin Xu

REVIEW 3 major objections 6 minor 11 references

AI2MMUM: AI-AI Oriented Multi-Modal Universal Model Leveraging Telecom Domain Large Model

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A single LLM-backed model, AI2MMUM, performs five physical-layer wireless tasks from one architecture, outperforming six ablated baselines on WAIR-D and DeepMIMO.

desk verdict A coherent universal-model architecture for wireless tasks, but the SOTA claim rests entirely on self-ablation comparisons. read the letter →

arxiv 2505.10003 v1 pith:CNJFDFD3 submitted 2025-05-15 cs.LG eess.SP

Tianyu Jiao , Zhuoran Xiao , Yihang Huang , Chenhui Ye , Yijia Feng , Liyu Cai , Jiang Chang , Fangkun Liu

show 4 more authors

Yin Xu Dazhi He Yunfeng Guan Wenjun Zhang

This is my paper · ORCID

classification cs.LGeess.SP

keywords AI2MMUM6Gmulti-modaluniversalmodelwirelesschannellargelanguageLoRAtaskinstructionsWAIR-D

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that one large-language-model-backed system, AI2MMUM, can perform five different physical-layer wireless tasks—direct positioning, LOS/NLOS identification, MIMO precoding, beam selection, and path loss prediction—from the same architecture, and outperforms six compared baselines on the WAIR-D and DeepMIMO datasets. The design freezes two contrastively pre-trained radio encoders, feeds their outputs together with task instructions into a telecom-domain LLM tuned with low-rank adaptation, and reads out each task through a lightweight linear head. The authors argue this matters because 6G networks would otherwise need a separate trained model for every task, which does not scale. The experiments are run on WAIR-D areas #00032 and #00247 and on DeepMIMO O1 BS#12, none of which the radio encoders saw during pre-training.

What carries the argument

The load-bearing mechanism is the pairing of frozen radio encoders (EPNN for physical environment, CFENN for CSI) with a trainable bridge: adapter layers map 128-dimensional radio embeddings into the LLM's token space, learnable prefix prompts (three tokens) plus fixed task keywords encode task identity, LoRA (rank 8) updates the attention query and key matrices, and a single linear head on the last token produces each task's output. The pre-trained encoders are the source of radio understanding; the LLM is the generalizing reasoner; the prompts and LoRA keep adaptation cheap.

What would settle it

Retrain the EPNN and CFENN encoders from scratch on the local data of WAIR-D areas #00032 and #00247 and DeepMIMO O1 BS#12, then compare their five-task performance with the frozen-encoder version; if the frozen version does not match or beat the local-trained version on the majority of tasks, the universal-representation claim fails. A reader could also inspect the pre-training split to confirm the test areas are truly absent from the 2.25M pairs.

Watch

Extended reading notes

Core claim

The paper's central claim is that task-agnostic radio representations, a task-instruction module with learnable prefix prompts, and a LoRA-tuned LLM backbone can be combined into one universal model that switches between regression and classification tasks without changing the backbone. On the five tested tasks, the full model outperforms six ablations: no learnable prompts, shared prompts, encoders trained from scratch, no LoRA, a randomly initialized LLM, and no LLM at all. The contribution is a new architecture and training recipe, not a new theory: the claim is that this particular combination transfers to unseen areas and to a second dataset.

Load-bearing premise

The whole scheme rests on the frozen radio encoders continuing to work in areas they never saw during pre-training; if that transfer fails, the model has no radio understanding to build on, and no checkpoint or code is provided to check it independently.

Editorial extensions

If this is right

If correct, one deployed model can replace separate task-specific wireless networks, because the same backbone handles all five tasks with only a change of instruction and head.
The frozen-encoder design means a new radio modality can be added by training a new adapter and LoRA, sparing the cost of re-training the whole model.
The ablation results imply that task-distinct instructions are not cosmetic: shared prompts degrade high-dimensional outputs such as precoding matrices and beam indices.
LoRA makes domain knowledge transferable to the LLM without unlearning its language abilities, since base weights stay frozen.

Reading between the lines

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

The paper's 'universal' claim is supported on only two WAIR-D areas and one DeepMIMO base station; a stronger test would sweep many more areas and base stations to map where the frozen encoders stop transferring.
Because LoRAs are modality-specific, the architecture suggests a natural scaling path to radar, LiDAR, and map inputs with a shared backbone; this is mentioned but not demonstrated in the experiments.
A reader could test whether the learnable prefix prompts encode stable task identities by training the same task from different random seeds and comparing the learned prompt embeddings; if they diverge, the prompts may be absorbing idiosyncratic noise rather than task semantics.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 reader's take is fair. The paper proposes a sensible and reasonably complete architecture for a 6G universal model—frozen contrastive radio encoders (EPNN/CFENN), a telecom-fine-tuned LLaMA-2 backbone with LoRA, fixed keyword plus learnable prefix prompts, and light task heads. The design choices are well motivated and the ablation set is thoughtful: it isolates the contribution of learnable prompts, task-distinct instructions, pre-trained encoders, LoRA, and the LLM itself. The results are internally consistent, with the full model generally outperforming its own ablations on all five tasks. That is real evidence that the components each carry some weight.

The problem is the word SOTA. The abstract claims state-of-the-art performance, but every comparison in Figs. 4–5 is against a variant of the proposed model. There is no independent baseline from the positioning, LOS/NLOS, precoding, beam-selection, or path-loss literature. The WM ablation (no LLM) is the only nod to a traditional method, and even that is not a tuned baseline. So the paper demonstrates internal gains over ablations, not superiority over existing methods. The stress-test note is correct, and this is the load-bearing weakness.

Other soft spots: no error bars or multiple seeds, so we cannot judge whether the small margins are meaningful. No code or checkpoints are released, and the frozen encoders come from the authors' own prior work [7,8] with no independent verification. The transfer to unseen areas and DeepMIMO is plausible but not independently checkable. The claim that the encoders exhibit 'scenario generalization' is taken as given from earlier papers. Minor: the paper has some rough language ('which flexibility and effectively') and the intro says it outperforms 'traditional non-LLM methods' without directly testing any published one.

On the positive side, the architecture is a plausible step toward a multi-task wireless foundation model, and the task-instruction design is worth trying. I would not cite the SOTA claim, but I would cite the architecture if I needed an example of an LLM-based universal model for wireless, provided the authors release code and add external baselines.

This deserves peer review, not desk rejection, because the core idea is serious and the ablations show self-consistency. But it needs major revision before acceptance.

Referee Report

3 major / 6 minor

Summary. The paper proposes AI2MMUM, a multi-modal universal model for 6G physical-layer tasks. The architecture combines frozen radio modality encoders (EPNN and CFENN from the authors' prior work), learnable prefix prompts plus fixed task keywords, a telecom LLM backbone fine-tuned with LoRA, and lightweight task-specific heads. The model is evaluated on five downstream tasks — direct positioning, LOS/NLOS identification, MIMO precoding, beam selection, and path loss prediction — using the WAIR-D and DeepMIMO datasets. The authors claim state-of-the-art performance, supported only by comparisons against six ablation variants of their own model.

Significance. If the system performs as described, the paper would provide a useful blueprint for an LLM-centric air-interface universal model: the combination of frozen pre-trained radio encoders, adapters, LoRA, and task-specific heads is coherent and the ablation design is internally consistent. The use of held-out WAIR-D areas and DeepMIMO O1 BS#12, which were not used in encoder pre-training, is a positive feature, as is the breadth of tasks considered. However, the headline claim of state-of-the-art performance is not supported by the experiments, because all compared 'benchmarks' are ablations of the proposed model rather than independent published baselines. The absence of error bars or multiple-seed runs further weakens the quantitative claims. Reproducibility is also limited because the encoder checkpoints and the telecom LLM are not released.

major comments (3)

[Abstract and §IV-C] The abstract claims that AI2MMUM 'achieves SOTA performance' on the five downstream tasks, but the experiments do not support a comparative state-of-the-art claim. The six methods named 'benchmarks' in §IV-B — FP, SP, TE/TC, WL, RL, and WM — are all ablations of the proposed architecture: removing learnable prompts, sharing prompts, training encoders from scratch, removing LoRA, randomizing the LLM, or removing the LLM entirely. None is an independent task-specific baseline from the literature. Figures 4 and 5 therefore demonstrate only that the full pipeline outperforms its own ablated variants, which is the expected outcome of an ablation study. To support the SOTA claim, the authors should add comparisons with published methods for each of the five tasks (or a subset with strong published baselines), or they should revise the abstract and conclusion to claim only that the proposed model outperforms its ablation variants.
[§IV-B and Figs. 4-5] All experimental results are reported as single point estimates with no error bars, no variance information, and no indication of the number of random seeds. Several of the observed margins are very small — for example, LOS/NLOS accuracy differences in Fig. 4(c) and (d) appear to be on the order of 0.1 to 0.2 percentage points, and path-loss RMSE differences in Fig. 5(b) are fractions of a decibel — so the reported gains may be within run-to-run noise. The paper should report mean and standard deviation over repeated training runs, and ideally include a significance test or confidence intervals for the key comparisons.
[§III-A and §IV-A] The central premise of the design is that the frozen EPNN and CFENN encoders, pre-trained in prior work [7], [8], provide universal and scenario-general representations that transfer to unseen areas. However, neither the encoders' weights nor the pre-training checkpoints are released, and the prior works are not described in sufficient detail for the reader to gauge the transfer risk. The authors should either release the encoder checkpoints (or provide a clear access mechanism) or include a self-contained transfer analysis, such as a zero-shot evaluation of the frozen encoders on the test areas independent of the final AI2MMUM pipeline, so that the claim of universal representation can be verified.

minor comments (6)

[Introduction] The phrase 'which flexibility and effectively perform various physical layer tasks' should be corrected to 'which can flexibly and effectively perform'.
[Abstract and throughout] The spacing in 'W AIR-D' is inconsistent; the dataset name should appear as 'WAIR-D'. Also, 'SOTA' should be spelled out at first use or defined in a footnote.
[Figs. 4-5] The figure legends use abbreviations FP, SP, TE/TC, WL, RL, and WM that are defined only in the text; the figure captions should include a brief note explaining these abbreviations, and the font size of the legends and axis labels should be increased.
[Table I] The abbreviations WC and PE are used in the table but defined only in the table title; please include the full names in the table caption or in a footnote.
[Conclusion] The conclusion states that AI2MMUM 'outperforms traditional non-LLM methods', but the only non-LLM method in the experiments is the WM ablation. This wording should be softened to match the evidence, e.g., 'outperforms the traditional non-LLM baseline used in our ablation studies'.
[Section I and II] The notation is mostly clear, but the term 'AI2MMUM' is typeset inconsistently as 'AI 2MMUM', 'AI2MMUM', and 'AI 2MMUM' across the paper; please standardize the spelling.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity found: the core pipeline is evaluated on held-out areas and independent datasets, though the SOTA claim rests on self-ablation comparisons rather than external baselines.

full rationale

The paper's derivation chain is not circular. The frozen EPNN/CFENN encoders were pre-trained in the authors' prior work on WAIR-D areas #01001 to #10000, while the downstream evaluations are conducted on the previously unseen WAIR-D areas #00032 and #00247 and on DeepMIMO O1 BS#12, so the reported results are not fitted to the pre-training inputs by construction. The telecom LLM backbone is likewise adopted from prior work [11] and is not re-derived from the downstream task labels. No equation in the paper defines a predicted output in terms of the fitted parameter being claimed as a prediction; the task heads, LoRA adapters, and prefix prompts are trained and evaluated under standard supervised settings. The main weakness is that the 'SOTA performance' claim in the abstract is supported only by comparisons against the paper's own ablations (FP, SP, TE/TC, WL, RL, WM), not against independent published baselines. That is an evidence and comparison-design limitation, not circularity: the full model could in principle lose to all ablations, so the comparisons are not forced to succeed by construction. The self-citations to [7], [8], and [11] are load-bearing components of the architecture, but their transferability is tested on data disjoint from their pre-training, which provides independent empirical content. Overall, no circular step rises to the level of Eq. X = Eq. Y by construction or fitted parameter renamed as prediction, and the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard ML assumptions plus a few hand-chosen hyperparameters and the transferability of the authors' own pre-trained encoders and telecom LLM. No new physical entities are introduced. The most consequential unpriced assumption is that the frozen EPNN/CFENN representations remain universal on unseen scenarios.

free parameters (3)

LoRA rank r = 8
Hand-chosen rank for low-rank adaptation; directly sets the 8.4M added parameters and affects the model's capacity to absorb wireless knowledge.
Learnable prefix prompt length = 3 tokens
Design choice for task instruction; Section IV-A states 'the learnable prefix prompts occupy three tokens'.
Adapter output dimension = 4096
Linear layer from 128-dim encoder output to 4096-dim LLM embedding, chosen to match the backbone dimension.

assumptions (4)

domain assumption Pre-trained EPNN/CFENN encoders from prior work [7][8] provide universal, task-agnostic radio representations.
Section III-A freezes these encoders and transfers them to unseen WAIR-D and DeepMIMO data; the paper does not independently validate their universality here.
domain assumption The telecom LLM [11], retrained from LLaMA2-7B on a telecom corpus, retains general language abilities while being fine-tuned with LoRA.
Section IV-A and the WL/RL ablations rely on this; no checkpoint is provided.
domain assumption The geometric ULA/OFDM channel model in Eqs. (1)-(3) is a valid representation of the WAIR-D and DeepMIMO channels.
The problem formulation uses this model, though the datasets are generated by simulation or ray-tracing; it is a modeling assumption rather than a tested claim.
domain assumption Task keywords such as 'position', 'LOS status', 'precoding', 'beam selection', and 'path loss' carry unambiguous task semantics for the LLM.
The Task Instruction Module in Section III-B relies on semantic consistency of fixed keywords; the SP ablation shows prompts matter but keyword sufficiency is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AI2MMUM: AI-AI Oriented Multi-Modal Universal Model Leveraging Telecom Domain Large Model." pith.science (2026). https://pith.science/paper/CNJFDFD3

@misc{pith2026250510003,
  author       = {Pith},
  title        = {Pith review of: AI2MMUM: AI-AI Oriented Multi-Modal Universal Model Leveraging Telecom Domain Large Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CNJFDFD3}},
  note         = {Machine review of arXiv:2505.10003}
}

read the original abstract

Designing a 6G-oriented universal model capable of processing multi-modal data and executing diverse air interface tasks has emerged as a common goal in future wireless systems. Building on our prior work in communication multi-modal alignment and telecom large language model (LLM), we propose a scalable, task-aware artificial intelligence-air interface multi-modal universal model (AI2MMUM), which flexibility and effectively perform various physical layer tasks according to subtle task instructions. The LLM backbone provides robust contextual comprehension and generalization capabilities, while a fine-tuning approach is adopted to incorporate domain-specific knowledge. To enhance task adaptability, task instructions consist of fixed task keywords and learnable, implicit prefix prompts. Frozen radio modality encoders extract universal representations and adapter layers subsequently bridge radio and language modalities. Moreover, lightweight task-specific heads are designed to directly output task objectives. Comprehensive evaluations demonstrate that AI2MMUM achieves SOTA performance across five representative physical environment/wireless channel-based downstream tasks using the WAIR-D and DeepMIMO datasets.

Figures

Figures reproduced from arXiv: 2505.10003 by the authors.

**Figure 1.** The 6G-oriented AI2MMUM capable of processing wireless multi-modal data and performing various air interface tasks. our method effectively extracts task-related features based on instructions and outperforms traditional non-LLM methods and models lacking our innovations across multiple sub-tasks. II. CHANNEL MODEL AND PROBLEM STATEMENT Taking the wireless channel modality in multi-modal data as an example, we consid… view at source ↗

**Figure 2.** Network structure of the proposed 6G-oriented, scalable, and task-aware AI [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Framework for communication multi-modal alignment. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The performance of our proposed method and six benchmarks across [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: The performance of our proposed method and six benchmarks across [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 8 canonical work pages

[7]

6G-oriented CSI-based multi-modal pre-training and down- stream task adaptation paradigm,

T. Jiao et al., “6G-oriented CSI-based multi-modal pre-training and down- stream task adaptation paradigm,” in2024 IEEE International Conference on Communications Workshops (ICC Workshops), Denver, CO, USA, 2024, pp. 1389-1394

work page 2024
[8]

Addressing the curse of scenario and task generalization in AI-6G: A multi-modal paradigm,

T. Jiao et al., “Addressing the curse of scenario and task generalization in AI-6G: A multi-modal paradigm,” inIEEE Transactions on Wireless Communications, early access, 2025

work page 2025
[1]

Big AI models for 6G wireless networks: Opportunities, challenges, and research directions,

Z. Chen, Z. Zhang, and Z. Yang, “Big AI models for 6G wireless networks: Opportunities, challenges, and research directions,” inIEEE Wireless Communications, vol. 31, no. 5, pp. 164-172, October 2024

work page 2024
[2]

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?,” inIEEE Communications Magazine, vol. 62, no. 11, pp. 84-90, November 2024

work page 2024
[3]

Large multi-modal models (LMMs) as universal foundation models for AI-native wireless systems,

S. Xu, C. K. Thomas, O. Hashash, N. Muralidhar, W. Saad, and N. Ra- makrishnan, “Large multi-modal models (LMMs) as universal foundation models for AI-native wireless systems,” inIEEE Network, vol. 38, no. 5, pp. 10-20, Sept. 2024

work page 2024
[4]

MAPLM: A real-world large-scale vision-language bench- mark for map and traffic scene understanding,

X. Cao et al., “MAPLM: A real-world large-scale vision-language bench- mark for map and traffic scene understanding,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 21819-21830

work page 2024
[5]

Talk2Radar: Bridging natural language with 4D mmWave radar for 3D referring expression comprehension,

R. Guan et al., “Talk2Radar: Bridging natural language with 4D mmWave radar for 3D referring expression comprehension,” 2024, arXiv:2405.12821

arXiv 2024
[6]

NetLLM: Adapting large language models for networking,

D. Wu et al., “NetLLM: Adapting large language models for networking,” inProceedings of the ACM SIGCOMM 2024 Conference, 2024, pp. 661- 678

work page 2024

Show all 11 references

[9]

W AIR-D: Wireless AI research dataset,

Y . Huangfu et al., “W AIR-D: Wireless AI research dataset,” 2022, arXiv:2212.02159

2022 arXiv
[10]

DeepMIMO: A generic deep learning dataset for mil- limeter wave and massive MIMO applications,

A. Alkhateeb, “DeepMIMO: A generic deep learning dataset for mil- limeter wave and massive MIMO applications,” 2019,arXiv:1902.06435

2019 arXiv
[11]

LLM agents as 6G orchestrator: A paradigm for task- oriented physical-layer automation,

Z. Xiao et al., “LLM agents as 6G orchestrator: A paradigm for task- oriented physical-layer automation,” in2024 IEEE Globecom Workshops (GC Wkshps), Cape Town, South Africa, 2024

2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [7]

6G-oriented CSI-based multi-modal pre-training and down- stream task adaptation paradigm,

T. Jiao et al., “6G-oriented CSI-based multi-modal pre-training and down- stream task adaptation paradigm,” in2024 IEEE International Conference on Communications Workshops (ICC Workshops), Denver, CO, USA, 2024, pp. 1389-1394

work page 2024

[2] [8]

Addressing the curse of scenario and task generalization in AI-6G: A multi-modal paradigm,

T. Jiao et al., “Addressing the curse of scenario and task generalization in AI-6G: A multi-modal paradigm,” inIEEE Transactions on Wireless Communications, early access, 2025

work page 2025

[3] [1]

Big AI models for 6G wireless networks: Opportunities, challenges, and research directions,

Z. Chen, Z. Zhang, and Z. Yang, “Big AI models for 6G wireless networks: Opportunities, challenges, and research directions,” inIEEE Wireless Communications, vol. 31, no. 5, pp. 164-172, October 2024

work page 2024

[4] [2]

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?,” inIEEE Communications Magazine, vol. 62, no. 11, pp. 84-90, November 2024

work page 2024

[5] [3]

Large multi-modal models (LMMs) as universal foundation models for AI-native wireless systems,

S. Xu, C. K. Thomas, O. Hashash, N. Muralidhar, W. Saad, and N. Ra- makrishnan, “Large multi-modal models (LMMs) as universal foundation models for AI-native wireless systems,” inIEEE Network, vol. 38, no. 5, pp. 10-20, Sept. 2024

work page 2024

[6] [4]

MAPLM: A real-world large-scale vision-language bench- mark for map and traffic scene understanding,

X. Cao et al., “MAPLM: A real-world large-scale vision-language bench- mark for map and traffic scene understanding,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 21819-21830

work page 2024

[7] [5]

Talk2Radar: Bridging natural language with 4D mmWave radar for 3D referring expression comprehension,

R. Guan et al., “Talk2Radar: Bridging natural language with 4D mmWave radar for 3D referring expression comprehension,” 2024, arXiv:2405.12821

arXiv 2024

[8] [6]

NetLLM: Adapting large language models for networking,

D. Wu et al., “NetLLM: Adapting large language models for networking,” inProceedings of the ACM SIGCOMM 2024 Conference, 2024, pp. 661- 678

work page 2024