{"id":"fa91ed2e-6d24-41a8-b864-f84c30d39852","arxiv_id":"2501.12983","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LLM4WM uses MoE-LoRA fine-tuning of a pre-trained LLM to jointly perform six wireless channel tasks, outperforming single-task baselines on simulated data.","lead":"A framework called LLM4WM fine-tunes a small GPT-2 model with mixture-of-experts adapters to handle six wireless channel tasks at once, from channel estimation to beamforming. It reports better average performance than single-task models and existing LLM-based baselines on simulated QuaDRiGa data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The full-sample claim survives, but the abstract's few-shot superiority is never actually tested; Section V-B2's 10%-data transfer is not a few-shot regime.","rationale":"I read the paper as a method-advance claim: one MoE-LoRA fine-tuned GPT-2 jointly handles six channel tasks and beats non-learning, small-model, and single-task baselines in full-sample and limited-data settings. The strongest evidence is Table IV, where LLM4WM is best or tied on all six tasks, and Table VI, where ablations degrade performance. I credit the parameter-efficiency numbers in Table VII and the expert-weight correlation analysis as useful supporting evidence. The reader's identified weakest assumption, Eq. (28)'s unnormalized average, is a valid statistical criticism: NMSE, (1-Acc), and MAE are on incompatible scales, so the single Avg. number is not meaningful. However, it is not the most load-bearing flaw because the full-sample ranking does not rely on the average: LLM4WM wins per-task. The same cannot be said for the abstract's 'few-shot evaluations.' The only limited-data experiment is Section V-B.2, which transfers on 10% of the RMa data, a low-resource setting that is not the few-shot regime promised, and no sample-count sweep is given. Since few-shot transfer is a core motivation for using an LLM at all, the conclusion 'outperforms existing methodologies in both full-sample and few-shot evaluations' is currently an overstatement. This is addressable with a conventional few-shot benchmark and normalized aggregation; hence the verdict remains conditional rather than reject or accept.","tokens_in":15384,"tokens_out":8497,"duration_ms":81855,"concrete_test":"Run a proper few-shot evaluation: for N in {1, 5, 10, 32} labeled samples per task (or per beam index), fine-tune LLM4WM, LLM4WM(s), LLM4CP, and the best small baseline on identical UMa and RMa splits, with at least 5 seeds. Report per-task metrics and a normalized aggregate (per-task rank or z-score) instead of Eq. (28). If LLM4WM is not consistently better than single-task fine-tuning and small baselines at low N, the abstract's few-shot superiority claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's Eq. (28) concern is real but not the crux: in Table IV, LLM4WM is best or tied on every per-task metric, so the full-sample superiority claim does not depend on the unnormalized Avg. column. The load-bearing problem is the abstract's promise of 'full-sample and few-shot evaluations.' Section V-B-2 reports a transfer experiment using only 10% of the RMa dataset. Given the 20,000-sample corpus in Section V-A-1, that is orders of magnitude more than a few-shot regime, and no N=1/5/10/32 learning curve or per-task sample count is reported. The only 'few-shot' mention in the Introduction refers to the LLM4CP baseline, not to the proposed evaluation. Because few-shot generalization is a distinct claimed advantage of MoE-LoRA plus a pretrained LLM, the evidence for that part of the central claim is missing. Moreover, in the RMa row of Table V, LLM4WM does not win every task (e.g., PE: WiT 0.220 vs 0.285), so the Avg. column computed by the unprincipled Eq. (28) is the only support for the overall transfer claim; the two issues compound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LLM4WM, a multi-task learning framework that fine-tunes a pretrained GPT-2 with Mixture-of-Experts Low-Rank Adaptation (MoE-LoRA) for six wireless channel-associated tasks: channel estimation, temporal and frequency channel prediction, sub-6G-assisted mmWave beamforming, distance estimation, and path loss estimation. The framework introduces task-specific preprocessing, input/output adapters to bridge the wireless feature space and the LLM semantic space, and dedicated output heads. Training is performed in two stages with DWA-based loss weighting. Experiments on QuaDRiGa-generated UMa/RMa channel data compare LLM4WM with traditional methods, single-task small models, multi-task small models, and single-task LLM baselines, covering full-sample performance, transfer/generalization, hyperparameter sensitivity, ablations, and efficiency. The paper claims that LLM4WM outperforms existing methodologies in both full-sample and few-shot evaluations.","tokens_in":1670,"tokens_out":1896,"duration_ms":75772,"significance":"If the claims are substantiated, the paper would demonstrate that a single frozen LLM with MoE-LoRA can jointly model several diverse wireless channel tasks with only about 1.13 M trainable parameters and transfer to new scenarios, a potentially useful step toward wireless foundation models. The strengths of the manuscript include a coherent task taxonomy, a broad set of baselines, ablations that isolate the contributions of the adapters and the backbone LLM, an expert-weight correlation analysis, and a parameter-efficiency comparison. However, the headline few-shot claim is not tested by the reported experiments, and the aggregate metric used for overall comparisons is not well-founded; these issues currently limit the strength of the paper's central claims.","major_comments":[{"comment":"The average metric in Eq. (28) is not a valid scalar for comparing methods across tasks because it sums quantities with different units and scales: three NMSE terms, a misclassification rate, and an MAE. This makes the Avg. column in Tables IV and V an arbitrary weighted combination, and any conclusion drawn from that column is not interpretable. The issue is load-bearing for the transfer claim in Table V: in the RMa row, LLM4WM is not the best method on BF (0.413 vs. CNN's 0.215) or PE (0.285 vs. WiT's 0.220), so the statement that LLM4WM consistently outperforms rests entirely on the unprincipled average. In Table IV the per-task comparisons are mostly favorable to LLM4WM, which is a point in the paper's favor, but the aggregate metric should be replaced with a normalized or standardized combination, or the claims should be based on per-task results with appropriate statistical testing.","section":"Section V-A-4, Eq. (28)"},{"comment":"The abstract's claim that LLM4WM is evaluated in few-shot settings is not supported by the experiments. The only few-shot statement in the Introduction concerns the LLM4CP baseline, not LLM4WM. The generalization experiment in Section V-B-2 uses only 10% of the RMa dataset to transfer a UMa-trained model; given the stated corpus of 20,000 samples, this is on the order of thousands of samples, which is not a few-shot regime, and no learning curves at N=1, 5, 10, or 32 or per-task sample counts are reported. Because few-shot generalization is one of the two headline claims of the paper, this missing evidence is a load-bearing gap. The authors should either add genuine few-shot experiments or revise the abstract and conclusion to remove the few-shot claim.","section":"Abstract and Section V-B-2"},{"comment":"The paper reports single performance numbers without error bars, confidence intervals, or multiple-seed results, yet several task-level differences are small: in Table IV, CE NMSE is 0.103 for LLM4WM versus 0.108 for LLM4CP, and CP NMSE is 0.106 for both LLM4WM and two baselines. Without uncertainty quantification, the claimed superiority over the closest baselines cannot be assessed, especially in the transfer experiments in Table V, where the task-level ranking differs across methods. This is a central issue because the paper's overall claim is that LLM4WM outperforms existing methodologies.","section":"Tables IV and V; Section V-B-1"},{"comment":"The ablation study reports only the unnormalized Average Loss defined in Eq. (28), so the loss-increase ratios inherit the arbitrary weighting problem described above. Per-task ablation numbers are needed to substantiate the conclusion that removing the backbone LLM has a significantly larger impact than removing the adapters. Additionally, the text says Results in Table V show... but the ablation results appear in Table VI; this cross-reference should be corrected.","section":"Section V-B-4 and Table VI"}],"minor_comments":[{"comment":"The notation for pilot spacing and frequency-domain slicing is ambiguous; please define npilot explicitly and use a consistent slice notation throughout.","section":"Section III-A, Eqs. (10a) and (12a)"},{"comment":"The table heading Interference time should be Inference time; the same typo appears in the accompanying text of Section V-B-5.","section":"Table VII and Section V-B-5"},{"comment":"The sentence about the smallest version of GPT-2 with F = 768 feature dimension is unclear: please define F and NL, and state exactly which layers of GPT-2 are used.","section":"Section V-A-3"},{"comment":"The description of the transfer experiment does not state how the 10% RMa subset is selected (random, stratified by task, etc.) or how many samples per task are used; this information is needed to interpret the transfer results.","section":"Section V-B-2"},{"comment":"The expert-weight correlation analysis is based on two randomly selected MoE-LoRA layers; a quantitative summary across all layers would strengthen the claim that the gating network learns task-differentiated expert combinations.","section":"Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of eess.SP and the proposed direction is timely. The main concern is that the few-shot claim in the abstract is unsupported, and the aggregate metric in Eq. (28) is not a principled basis for the overall comparisons. These issues are fixable with additional experiments or revised claims. I did not find evidence of problematic citation practices; the self-citations are to directly related prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: This paper assembles a sensible multi-task LLM framework for wireless channel tasks—MoE-LoRA on a frozen GPT-2, per-task adapters, and task-specific pre/post processors applied to six tasks—and the full-sample results mostly support it. LLM4WM wins or ties most per-task metrics in Table IV, and the ablation table shows both the LLM backbone and the adapters contribute. The trainable parameter count (~1.13 M) is a practical plus. That part is worth a referee's time.\n\nThe soft spots are in the evaluation. Eq. (28)'s 'Avg.' metric sums four NMSE terms, a misclassification rate, and an MAE without normalization; the scales differ by orders of magnitude, so the single average is not interpretable. In Table IV the per-task results carry the conclusion, so the bad metric is mainly a presentation problem. But in the transfer experiment (Table V), LLM4WM doesn't win every task in the RMa row (WiT beats it on PE), so the Avg. column becomes load-bearing. Worse, the paper calls this a 'few-shot' evaluation. It is a transfer experiment using 10% of the RMa dataset—with a 20,000-sample base, that is orders of magnitude away from few-shot, and no N=1/5/10/32 curve or per-task sample counts are provided. The abstract's 'few-shot evaluations' claim is therefore unsupported. The paper also gives no error bars or multiple seeds, and the LLM4CP baseline adaptation is described only as 'LN Tuning on the same GPT-2,' which is thin for a comparison to prior work by the same group.\n\nNone of this sinks the framework. The per-task full-sample results, the expert-weight correlation analysis, and the ablations show a working method. The paper just overclaims on transfer and underreports uncertainty. A revision that adds seeds, normalizes or drops the Avg. metric, and either runs real few-shot experiments or drops the term would make the claims match the evidence.","headline":"Useful multi-task LLM framework for wireless, but the few-shot claim outruns the experiments.","tokens_in":16201,"tokens_out":3340,"would_cite":true,"duration_ms":31861,"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 single LLM fine-tuned with Mixture-of-Experts LoRA can jointly handle six wireless channel tasks and beat per-task models, especially in few-shot transfer.","keywords":["large language models","Mixture of Experts","Low-Rank Adaptation","multi-task learning","wireless channel","channel estimation","beam management","transfer learning"],"falsifier":"Recompute the average metric used in Tables IV-VI after normalizing each task's error by its own baseline scale (or by its spread across methods), then compare LLM4WM against the same baselines; if the ranking changes or the per-task margins shrink, the paper's headline overall comparison no longer holds.","tokens_in":15220,"feed_emoji":"📡","tokens_out":8362,"duration_ms":80245,"temperature":0.7,"pith_summary":"The paper proposes LLM4WM, a way to adapt a pre-trained language model into a single multi-task engine for six wireless channel problems: channel estimation, time-domain channel prediction, frequency-domain channel prediction, sub-6 GHz-assisted mmWave beam selection, distance estimation, and path-loss estimation. The central claim is that one LLM fine-tuned with Mixture-of-Experts Low-Rank Adaptation (MoE-LoRA), plus task-specific adapters at input and output, learns a shared channel representation that improves each task and transfers better to new scenarios than training per-task models. On a simulated dual-frequency (sub-6 GHz and mmWave) dataset the paper reports lower errors and higher accuracy than non-learning baselines, small single-task models, and single-task LLM fine-tuning, with particularly strong few-shot transfer results. If this holds, wireless systems could replace a collection of task-specific models with one parameter-efficient fine-tuned LLM that gains a new task by adding roughly a million trainable parameters.","feed_headline":"One LLM beats per-task models at six wireless channel jobs","feed_subtitle":"MoE-LoRA fine-tuning shares channel structure across estimation, prediction, beams, and ranging.","key_machinery":"The load-bearing mechanism is MoE-LoRA, a multi-task extension of Low-Rank Adaptation inserted into the LLM's feed-forward layers. Standard LoRA writes the updated weight as $W = W_0 + \\frac{\\alpha}{r}BA$ with two low-rank trainable matrices $A$ and $B$; MoE-LoRA instead maintains $N_e$ expert pairs and computes $W = W_0 + \\frac{\\alpha}{r}\\sum_{k=1}^{N_e}\\omega_k B_k A_k$, where a single-layer softmax gating network produces the per-task expert weights $\\omega_k$. This keeps the backbone frozen while letting tasks share some expert combinations and diverge on others, which is how the model balances common channel structure against task-specific features. Around this, the multi-task adapters align wireless feature tensors with the LLM's semantic space, and the task-specific preprocessors and heads translate raw channel matrices and output formats. The other named component is the Dynamic Weight Average loss, which re-weights each task's loss every epoch to avoid one task dominating training.","core_discovery":"On the paper's own terms, the discovery is that channel-associated wireless tasks are similar enough that a single large language model, kept mostly frozen, can serve as a shared backbone for all of them. LLM4WM preprocesses each task's channel data (tokenizing CSI for reconstruction tasks, applying a DFT for angle-sensitive tasks), aligns it to the LLM's hidden space through an input multi-task adapter, processes it through the frozen transformer whose feed-forward layers are updated only by low-rank expert combinations, and maps the output through a second adapter and a task-specific CNN or MLP head. The reported result is that this joint model beats every listed baseline on the per-task metrics, including NMSE for channel reconstruction and path loss, Top-1 accuracy for beam management, and MAE for distance estimation, and that the advantage grows in few-shot transfer from one scenario and frequency band to another. The authors also report that removing the LLM backbone hurts performance more than removing the adapters, which they read as evidence that the pre-trained model's general representations, not the task-specific modules alone, carry the multi-task benefit.","pith_inferences":["Editorial inference: the reported 'Avg.' metric in Eq. (28) sums raw NMSEs, a misclassification rate, and a MAE, which have different units and scales; a normalized or task-balanced average could change the overall ranking, so per-task numbers should be read before the average.","Editorial inference: the framework's success on simulated channels suggests testing it on measured channel data; real-world distribution shifts are usually larger than the scenario-to-scenario transfer tested here and would be the stricter test of the shared-representation claim.","Editorial inference: the expert-weight correlation structure could be reused as a tool for task clustering, letting an operator decide which new wireless tasks should be added to the same model and which should get their own expert group.","Editorial inference: nothing in the method is specific to wireless; the same adapter plus MoE-LoRA recipe could be tried on other non-linguistic signal domains with multiple correlated tasks, such as radar sensing or spectrum monitoring."],"forward_implications":["A single fine-tuned LLM can take over six channel-related tasks that today are handled by separate small models, reducing the number of models a base station must train and serve.","Adding a new task is cheap: the paper reports about 1.13 million trainable parameters added per task, so the framework scales to larger task sets with modest memory and training cost.","Few-shot deployment becomes practical: after fine-tuning on only 10 percent of data from a new scenario or band, LLM4WM retains more accuracy than per-task baselines, reducing the need to collect large labeled datasets on every deployment.","The gating network learns a task taxonomy on its own: expert-weight correlations are low for dissimilar tasks and higher for related ones, suggesting the model discovers which wireless tasks share structure without being told.","Ablations indicate the pre-trained LLM is doing the heavy lifting, so future versions should invest in better backbone models and routing rather than only in task-specific heads."],"supporting_citations":[{"why":"Supplies the single-task LLM channel-prediction baseline that LLM4WM extends and compares against.","marker":"[19]"},{"why":"Supplies the MoE-LoRA parameter-efficient multi-task fine-tuning method adapted for wireless tasks.","marker":"[21]"},{"why":"Provides the sparsely-gated Mixture-of-Experts layer that motivates the expert gating design.","marker":"[31]"},{"why":"Supplies the Dynamic Weight Average algorithm used to balance task losses during training.","marker":"[33]"},{"why":"Provides the channel generator used to create the dual-frequency sub-6 GHz and mmWave dataset.","marker":"[34]"},{"why":"Transformer-based channel predictor used as a baseline for the reconstruction tasks.","marker":"[23]"},{"why":"CNN-based uplink-to-downlink channel prediction baseline for the reconstruction tasks.","marker":"[24]"},{"why":"Cross-stitch multi-task network used as the multi-task small-model baseline.","marker":"[40]"}],"fun_headline_variants":["Single LLM tops per-task models on six wireless channel jobs","LLM4WM: one LLM, many wireless channel tasks","Frozen LLM backbone beats wireless task specialists","Shared LLM wins on wireless tasks even with few examples","MoE-LoRA lets one LLM handle all wireless channel tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported headline 'Avg.' score in Eq. (28) sums four normalized mean-square errors, a misclassification rate, and a mean absolute error without normalizing their scales, so the claimed overall superiority over baselines rests on this unweighted combination being a meaningful measure of performance.","fun_headline_variants_meta":{"raw":{"variants":["Single LLM tops per-task models on six wireless channel jobs","LLM4WM: one LLM, many wireless channel tasks","Frozen LLM backbone beats wireless task specialists","Shared LLM wins on wireless tasks even with few examples","MoE-LoRA lets one LLM handle all wireless channel tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000449,"raw_usage":{"total_tokens":2257,"prompt_tokens":927,"completion_tokens":1330,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":1245}},"tokens_in":543,"tokens_out":1330,"duration_ms":9923,"temperature":1.0,"reasoning_tokens":1245,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:33:17.721016+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the average metric used in Tables IV-VI after normalizing each task's error by its own baseline scale (or by its spread across methods), then compare LLM4WM against the same baselines; if the ranking changes or the per-task margins shrink, the paper's headline overall comparison no longer holds.","supporting_citations":[{"cited_title":"Cross-stitch net- works for multi-task learning,","cited_arxiv_id":null,"evidence_quote":"Cross-stitch multi-task network used as the multi-task small-model baseline."},{"cited_title":"Deep UL2DL: Data- Driven Channel Knowledge Transfer From Uplink to Downlink,","cited_arxiv_id":null,"evidence_quote":"CNN-based uplink-to-downlink channel prediction baseline for the reconstruction tasks."},{"cited_title":"LLM4CP: Adapting Large Language Models for Channel Prediction,","cited_arxiv_id":null,"evidence_quote":"Supplies the single-task LLM channel-prediction baseline that LLM4WM extends and compares against."},{"cited_title":"When MOE Meets LLMs: Parameter Efficient Fine-tuning for Multi-task Medical Applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the MoE-LoRA parameter-efficient multi-task fine-tuning method adapted for wireless tasks."},{"cited_title":"End-to-End Multi-Task Learning with Attention,","cited_arxiv_id":null,"evidence_quote":"Supplies the Dynamic Weight Average algorithm used to balance task losses during training."},{"cited_title":"QuaDRiGa: A 3- D Multi-Cell Channel Model With Time Evolution for Enabling Virtual Field Trials,","cited_arxiv_id":null,"evidence_quote":"Provides the channel generator used to create the dual-frequency sub-6 GHz and mmWave dataset."},{"cited_title":"Accurate Channel Prediction Based on Transformer: Making Mobility Negligible,","cited_arxiv_id":null,"evidence_quote":"Transformer-based channel predictor used as a baseline for the reconstruction tasks."}],"review_version":1}