{"id":"c822602d-76b1-4b79-a12d-e944db3692d7","arxiv_id":"2506.17067","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A proposed GPT-2-based framework for jointly classifying far versus near-field users and designing multiuser precoding in low-altitude drone networks, supported by a brief simulation and many open questions.","lead":"This article describes how large language models can be added to next-generation wireless base stations to tell nearby and distant drones apart and shape their signals. It is a position and framework paper with a small simulation showing the idea can beat older baselines, but no code or detailed results are provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central performance claim is unsupported: Section IV-B reports no numerical spectrum efficiencies, accuracies, or error bars, so 'excellent performance' and 'outperforms' cannot be checked even before assessing whether frozen GPT-2 features transfer to CSI.","rationale":"The reader's weakest assumption about frozen GPT-2 transfer is a plausible specific mechanism, but the more load-bearing issue is one level earlier: the manuscript provides no quantitative evidence from which that premise, or any performance claim, could be assessed. The central claim is empirical, so missing numbers, error bars, code, and baseline specifications directly block verification. I do not see an internal inconsistency in the near-field channel model or the beamforming formula; the beamforming structure cited from [12] is standard. The near-field/LLM framework is coherent as a proposal. The deployment limitations acknowledged in Section V-C (computational cost, battery life) do not undermine the claimed simulation performance. The novelty relative to [13] is unclear but is a secondary issue compared with the absence of measurable results. Because the reader's CONDITIONAL verdict already requires substantiation, my concern reinforces that verdict without moving it to REJECT: the paper could be made acceptable by adding a reproducible, quantitative case study with statistical significance testing. Thus I recommend UNCHANGED (i.e., keep CONDITIONAL).","tokens_in":7771,"tokens_out":6056,"duration_ms":67384,"concrete_test":"Require the authors to release the code, data, and trained model (or to provide a complete implementation specification), then reproduce Fig. 4 on the stated 8k/1k/1k split. Report numerical spectrum efficiency and classification accuracy with at least five random seeds and error bars for the proposed method and every baseline. If the LLM-based method does not exceed the CNN/Transformer baselines by a statistically significant margin (e.g., non-overlapping confidence intervals or a paired test), the central outperformance claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: the proposed LLM-based scheme 'demonstrates excellent performance under various parameter settings' and 'outperforms other deep learning-based methods' (Section IV-B, Fig. 4). The only evidence is a qualitative figure with no numerical values reported in the text, no error bars, no released code or data, and no specification of the CNN/Transformer baselines' architectures, parameter counts, or training budgets. Because both the precoding spectrum efficiency and the far/near classification accuracy are presented as curves without numbers, there is no way to determine whether the proposed method's advantage over CNN/Transformer is real, large, or statistically significant. This is more immediate than the transferability of frozen GPT-2 representations: even if a frozen GPT-2 can extract useful structure from complex CSI, the manuscript still fails to substantiate the claim that it does so. The unspecified 'Discretization' and 'Domain Transformation' in the preprocessing module (Section IV-A) could destroy the phase information that distinguishes near-field users, but this is secondary; the primary gap is the absence of any quantitative, reproducible evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes applying large language models (LLMs) to near-field communications in low-altitude economy (LAE) networks. It argues that LAE user equipment is naturally located in the near-field region of XL-MIMO base stations, benefiting from spherical-wave beamfocusing and the additional distance dimension. The authors identify signal processing complexity and far/near-field user classification as key challenges, and they propose an LLM-based framework with input encoders, a GPT-2 backbone with frozen self-attention/MLP layers, and output decoders for two tasks: distinguishing far- and near-field users and designing multi-user precoding. A case study is presented with simulation results claimed to show the proposed scheme outperforms CNN, Transformer, and conventional codebook-based schemes. The paper also discusses future directions such as multi-task and multi-modal LLMs for LAE.","tokens_in":8047,"tokens_out":2765,"duration_ms":29375,"significance":"If the central empirical claim were substantiated, the paper would demonstrate a novel application of pretrained LLMs to physical-layer wireless tasks, specifically near-field user classification and precoding in XL-MIMO for LAE. The main strength is the system-level framing: connecting near-field propagation, LAE mobility, and a single LLM backbone could, in principle, provide a unified solution to multiple tasks. The authors also correctly cite a known structure for optimal downlink beamforming to reduce the learning target from a full precoding matrix to a per-user power-allocation-like parameter vector. However, the paper does not currently provide any verifiable quantitative evidence for its headline claims: no numerical values, no error bars, no baseline specifications, and no ablation isolating the contribution of the pretrained LLM. Until such evidence is supplied, the significance remains potential rather than demonstrated.","major_comments":[{"comment":"The central claim that the proposed scheme 'demonstrates excellent performance under various parameter settings' and 'outperforms other deep learning-based methods' is not supported by any quantitative data in the manuscript. The text reports no numerical values for spectrum efficiency, classification accuracy, or comparison margins; no error bars or confidence intervals are given; and the baseline schemes (CNN, Transformer, NOMA, LDMA, SDMA) are not specified in terms of architecture, parameter count, or training budget. Without these numbers, the claimed superiority cannot be checked, and the reader cannot distinguish a genuine gain from random variation or favorable axis scaling.","section":"Section IV-B, Fig. 4"},{"comment":"The training recipe is critically under-specified. The paper does not define loss functions for the classification and precoding tasks, does not provide the channel model equations, does not describe how the 8,000/1,000/1,000 training/validation/testing samples are generated, and does not specify the 'Discretization' and 'Domain Transformation' steps in the pre-processing module. These details are essential for reproducibility and for assessing whether the preprocessing preserves the phase information that distinguishes near-field users from far-field users. The statement that only addition and layer-normalization layers are fine-tuned while self-attention and MLP layers are frozen is a key architectural choice, but the paper provides no ablation or feature-analysis evidence that frozen text-pretrained representations transfer to raw complex channel state information.","section":"Section IV-A, 'Pre-processing Module' and 'LLM Backbone With Finetuning'"},{"comment":"The optimal beamforming formula is cited from [12] rather than derived or adapted to the near-field channel model, and the mapping from the LLM output to the parameter vector λ and the power allocation vector p is not specified. The paper does not state the output projection dimensions, activation functions, how the power constraint is enforced, or how the λ parameters are converted into the final precoding vectors. Because the precoding task is a central component of the case study, this missing specification weakens the entire empirical section.","section":"Section IV-B2, Eq. (1)"}],"minor_comments":[{"comment":"The phrase 'Beside, compared with users on the ground' should be 'Besides, compared with users on the ground'.","section":"Section III-A"},{"comment":"The caption contains a typographical error: 'θ_tit = 5°' should be 'θ_tilt = 5°'. The variables in the plots ('Spectrum efficiency against ?', 'against ?N', 'against ?min') are not defined in the caption or the text.","section":"Fig. 4 caption"},{"comment":"The subsection titles 'Input encoders', 'Output decoders', and 'LLM Backbone With Finetuning' are inconsistently capitalized and use different formatting from 'Input Module' and 'Output Module'; please standardize them.","section":"Section IV-A, subsection titles"},{"comment":"Reference [11] lists the first author as 'R. K. Alec', which does not match the standard citation for the GPT-2 technical report (Alec Radford et al.). The authors should correct this reference.","section":"References"},{"comment":"The spacing in 'UA Vs' is inconsistent (sometimes 'UAVs', sometimes 'UA Vs'); please unify the notation. Also, 'fine-tuning' and 'finetuning' are used interchangeably; please choose one spelling.","section":"Throughout"},{"comment":"The discussion of the authors' prior work [13] is brief and does not clearly delineate the novelty of the proposed scheme relative to [13]. Since the proposed framework closely resembles the multi-task LLM network in [13], the authors should explicitly state what is new in this paper beyond that prior work.","section":"Section V-A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript closely mirrors the authors' prior multi-task LLM physical-layer network [13] and appears to be a magazine-style overview rather than a full research contribution. The absence of any quantitative experimental results is a serious editorial concern; the editor may wish to request a revised version with a proper numerical evaluation, full training and channel-model details, and a clear statement of novelty relative to [13]. If the authors cannot produce such results, the paper should be considered for rejection on the grounds that its central claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a tutorial-style paper with a plausible idea—use a fine-tuned GPT-2 to both classify far/near-field users and compute multiuser precoding in XL-MIMO for low-altitude drone networks. The new bit is the specific application, not the machinery: the framework is a direct adaptation of the authors' own multi-task LLM network [13], and the optimal precoding formula is taken from Xia et al. [12]. That's fine, but it means the paper's value has to come from the case study. And the case study is where it falls short.\n\nWhat it does well: the survey sections are competent. The near-field basics (spherical wave model, beamfocusing, increased DoF) are explained clearly, and the LAE context gives concrete motivation. The authors are honest about provenance—they cite [12] for the beamforming structure and mention [13] as prior multi-task work. The future directions are sensible, not padded.\n\nWhere it gets soft: the central performance claim is 'excellent performance' and 'outperforms other deep learning-based methods,' but the only evidence is Fig. 4, which contains no numbers, no error bars, and no baseline details. You don't know the CNN/Transformer architectures, parameter counts, or training budgets. The preprocessing steps (Discretization, Domain Transformation) are named but not specified—not a minor omission when phase is what distinguishes near-field users. The training recipe is also thin: no loss functions, no channel model equations, no data generation description. So the case study is more sketch than verification. The stress-test note is right: missing quantitative evaluation is the immediate problem, before you even reach the question of whether frozen GPT-2 features transfer to CSI.\n\nI disagree with the reader's circularity concern: no equation is derived from its own output; the precoding is cited and the LLM is trained. The real issue is under-specification, not circularity. Also, the contribution relative to [13] is incremental and not clearly delineated.\n\nWho this is for: people who want a bird's-eye view of LLM + near-field + LAE, or someone curious about the authors' direction. That's a legitimate audience. But as a technical paper it's not ready; the empirical claim cannot be checked.\n\nMy recommendation: if I were the editor of a magazine/tutorial venue, I'd send it to review and require a substantially expanded simulation section before acceptance. At a regular journal, I'd desk reject it. I'd bring it to reading group only to debate the hype around LLMs in physical-layer processing.","headline":"A readable tutorial that floats a plausible GPT-2-for-near-field idea, but the case study has no numbers and the contribution over [13] is thin.","tokens_in":8542,"tokens_out":3093,"would_cite":false,"duration_ms":31193,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A pretrained GPT-2, with most layers frozen, can be fine-tuned to distinguish far- and near-field users and compute multiuser precoding in downlink XL-MIMO for low-altitude drone networks.","keywords":["large language models","near-field communications","extremely large-scale MIMO","low-altitude economy","beamfocusing","multiuser precoding","far/near-field classification","GPT-2"],"falsifier":"Train the proposed architecture twice under identical data and hyperparameters, once with the published GPT-2 weights and once with the same architecture but randomly initialized transformer layers. If the pretrained and random versions perform equally on far/near-field classification accuracy and achieved spectrum efficiency, the text pretraining is not load-bearing and the reported gains would reduce to the task-specific encoders and fine-tuned layers.","tokens_in":7566,"feed_emoji":"📡","tokens_out":8511,"duration_ms":73549,"temperature":0.7,"pith_summary":"This article proposes using a large language model as the processing core of a downlink extremely large-scale MIMO system serving low-altitude drones. The central claim is that one pretrained GPT-2 backbone, with its self-attention and MLP layers kept frozen, can be fine-tuned to solve two tasks at once: telling far-field from near-field users and computing the multiuser precoding matrix. The authors report a case study in which the proposed scheme reaches high classification accuracy and higher spectrum efficiency than codebook-based and deep-learning baselines across varied system parameters. If the claim holds, text-pretrained transformers would provide a single reusable engine for physical-layer jobs in near-field UAV networks.","feed_headline":"A frozen GPT-2 can classify and beam to near-field UAV users","feed_subtitle":"Lightly tuned GPT-2 sorts far- and near-field users and sets multiuser precoding, beating deep-learning baselines in simulation.","key_machinery":"The central object is the six-module LLM-based framework, with the frozen GPT-2 backbone (dimension $d=768$) as its core. The key mechanism is the split between frozen and trainable parts: self-attention and MLP layers are frozen to retain universal knowledge, while only the residual 'addition' and layer-normalization layers are fine-tuned for the wireless tasks. For precoding, the paper uses the proven optimal beamforming form $w_k^*=(\\mathbf{I}_N+\\sum_{k=1}^K \\frac{\\lambda_k}{\\sigma^2} \\mathbf{h}_k \\mathbf{h}_k^H)^{-1} \\mathbf{h}_k / \\|(\\mathbf{I}_N+\\sum_{k=1}^K \\frac{\\lambda_k}{\\sigma^2} \\mathbf{h}_k \\mathbf{h}_k^H)^{-1} \\mathbf{h}_k\\|$, so the network learns only the $K$ scalars $\\lambda_k$ and the power allocation $\\mathbf{p}$ instead of the whole matrix $\\mathbf{W}$; this dimensionality reduction is what makes the precoding task tractable for a language-model backbone.","core_discovery":"The paper's discovery claim is that GPT-2, a text-trained transformer, transfers to physical-layer signal processing after only residual-connection and layer-normalization layers are fine-tuned. In the proposed pipeline, raw complex channel matrices are concatenated, normalized, discretized, and encoded, then passed through the frozen GPT-2 backbone; output decoders map the internal features to a far/near-field label per user and to the scalar power-allocation parameters $\\lambda_k$ that determine the precoding vectors via $w_k^* = (\\mathbf{I}_N + \\sum_{k} \\frac{\\lambda_k}{\\sigma^2} \\mathbf{h}_k \\mathbf{h}_k^H)^{-1} \\mathbf{h}_k / \\|(\\mathbf{I}_N + \\sum_{k} \\frac{\\lambda_k}{\\sigma^2} \\mathbf{h}_k \\mathbf{h}_k^H)^{-1} \\mathbf{h}_k\\|$. The case study reports that this scheme demonstrates excellent performance under various parameter settings and outperforms other deep learning-based methods on multiuser spectrum efficiency.","pith_inferences":["If the transfer result is real, text-pretrained transformers could become a common backbone for other complex-valued channel tasks, such as channel prediction or integrated sensing, where the input is not language at all.","A direct test of the mechanism would ablate pretraining: replace GPT-2 weights with random weights and compare, a comparison the paper does not report.","The proposed pipeline could likely work with much smaller language models, since the wireless task uses only 256-antenna channels; scaling studies would reveal whether the LLM size is actually buying performance or just capacity.","Real deployment would need to weigh the GPT-2 backbone's inference cost against the claimed gains on battery-limited UAVs, an issue the paper lists as an open challenge."],"forward_implications":["A single frozen GPT-2 backbone can serve multiple physical-layer tasks at once, so user classification and precoding need not be separate specialist networks.","Near-field beamfocusing and location division multiple access become compatible with low-altitude drone networks, since the model can place users in the distance domain instead of treating all links as planar-wave.","Because only residual and layer-norm layers are fine-tuned, adapting the model to a new wireless scenario costs less than retraining a full transformer.","The same LLM-based scheme should extend to channel prediction and user localization, tasks the authors list as future outputs of the framework.","LLM-based methods can surpass codebook-based beamforming and CNN/Transformer baselines in spectrum efficiency as the antenna array grows."],"supporting_citations":[{"why":"Supplies the GPT-2 backbone whose frozen representation is the core of the scheme.","marker":"[11]"},{"why":"Gives the optimal downlink beamforming structure, letting the network learn only the scalar parameters.","marker":"[12]"},{"why":"Provides the multi-task physical-layer LLM framework that this article adapts to near-field LAE.","marker":"[13]"},{"why":"Defines near-field XL-MIMO fundamentals and the spherical-wave model used throughout.","marker":"[5]"},{"why":"Motivates the far/near-field classification problem from the interference perspective.","marker":"[10]"},{"why":"Supplies the near-field beamfocusing result that the low-altitude network exploits.","marker":"[6]"}],"fun_headline_variants":["Lightly tuned GPT-2 sorts near-field users and sets precoding","Frozen GPT-2 transfers to near-field beamforming, beating DL baselines","GPT-2, fine-tuned on residuals, classifies and beams to UAVs","LLM meets near-field: GPT-2 with minimal tuning outperforms DL","Frozen GPT-2, only residual/layernorm tuned, steers near-field MU-MIMO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire gain rests on the premise that the knowledge stored in GPT-2's frozen text-trained weights transfers to complex-valued wireless channel matrices after generic preprocessing, so that fine-tuning only residual and layer-norm layers is enough to adapt the model to far/near-field classification and precoding.","fun_headline_variants_meta":{"raw":{"variants":["Lightly tuned GPT-2 sorts near-field users and sets precoding","Frozen GPT-2 transfers to near-field beamforming, beating DL baselines","GPT-2, fine-tuned on residuals, classifies and beams to UAVs","LLM meets near-field: GPT-2 with minimal tuning outperforms DL","Frozen GPT-2, only residual/layernorm tuned, steers near-field MU-MIMO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000748,"raw_usage":{"total_tokens":3369,"prompt_tokens":1019,"completion_tokens":2350,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":2241}},"tokens_in":635,"tokens_out":2350,"duration_ms":16580,"temperature":1.0,"reasoning_tokens":2241,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:12:41.409634+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the proposed architecture twice under identical data and hyperparameters, once with the published GPT-2 weights and once with the same architecture but randomly initialized transformer layers. If the pretrained and random versions perform equally on far/near-field classification accuracy and achieved spectrum efficiency, the text pretraining is not load-bearing and the reported gains would reduce to the task-specific encoders and fine-tuned layers.","supporting_citations":[{"cited_title":"Language models are unsupervised multitask learners,","cited_arxiv_id":null,"evidence_quote":"Supplies the GPT-2 backbone whose frozen representation is the core of the scheme."},{"cited_title":"A deep learning framework for optimization of MISO downlink beam- forming,","cited_arxiv_id":null,"evidence_quote":"Gives the optimal downlink beamforming structure, letting the network learn only the scalar parameters."},{"cited_title":"A tutorial on near-field XL-MIMO communications towards 6G,","cited_arxiv_id":null,"evidence_quote":"Defines near-field XL-MIMO fundamentals and the spherical-wave model used throughout."},{"cited_title":"Mixed near-and far- field communications for extremely large-scale array: An interference perspective,","cited_arxiv_id":null,"evidence_quote":"Motivates the far/near-field classification problem from the interference perspective."}],"review_version":2}