{"id":"99191ba5-466a-402a-9501-cf1315afcd16","arxiv_id":"2506.07647","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper proposes a systematic classification and two roadmaps for using foundation models (LLMs and wireless foundation models) to design Synesthesia of Machines systems for 6G, with preliminary case-study evidence of gains over task-specific models.","lead":"This paper organizes how foundation models, such as large language models and specialized wireless models, can be applied to future 6G networks that combine communication and multi-modal sensing. It proposes two design roadmaps and reports preliminary simulations showing that foundation models outperform task-specific models on channel prediction, path loss generation, and cooperative perception tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LLM roadmap's core premise—language-pretrained GPT-2 transfers physical structure to wireless tasks—is asserted, not tested; a randomized-initialization ablation would settle it.","rationale":"The reader's weakest-assumption identification is accurate: the LLM roadmap is the part of the paper that genuinely requires cross-domain transfer, and it is the least empirically supported. The wireless foundation model roadmap is more defensible: WiFo and WiPo pre-train on in-domain data (QuaDRiGa CSI, ImageNet, CsiNet-Outdoor, ShapeNet), and their gains over from-scratch baselines are the expected outcome of pre-training, even if hyperparameters and error bars are missing. The LLM case studies instead ask a frozen language model to encode physical structure. Section III.B describes bridging the 'significant gap between the natural language domain and the multi-modal information domain' by feature-level fusion and tokenization, but this bridging is exactly where the assumption lives; no experiment shows that the frozen transformer blocks contribute beyond what a randomly initialized transformer with the same fine-tuning could provide. Several case studies are re-presentations of prior work (WiFo [104], LLM4WM [72], LLM4SG [120]), which is legitimate but means the new empirical content is limited. The absence of error bars and code is a secondary concern, not the load-bearing one. A randomized-initialization/embedding-permutation ablation is cheap and directly settles whether language pretraining is the source of the reported gains. Until then, CONDITIONAL is the right verdict, and the reader's assessment does not need to change.","tokens_in":29329,"tokens_out":6267,"duration_ms":79333,"concrete_test":"Run the LLM4PG, LLM4SG, and LLM4WM pipelines with three GPT-2 variants under identical data, adaptation (LN Tuning/MoE-LoRA), and training budget: (1) the published pretrained GPT-2; (2) the same architecture with randomly initialized weights; (3) pretrained GPT-2 with token/positional embeddings randomly permuted, preserving architectural capacity and parameter count while destroying linguistic structure. Repeat with at least 5 seeds and report mean±std. If (2) or (3) performs within noise of (1) on the reported metrics (path-loss map error, scatterer accuracy, NMSE), the 'general knowledge transfer' premise is unsupported and the LLM roadmap's advantage vanishes. If (1) is clearly and consistently better, the transfer assumption survives this test. If LN-only tuning makes random-init untrainable, repeat with full fine-tuning for all variants.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that FMs systematically address SoM's four challenges depends on both roadmaps, and Roadmap 1 rests on the premise that GPT-2's frozen text-pretrained transformer—with only LayerNorm parameters trainable in LLM4PG/LLM4SG (§III.B Step 2) or MoE-LoRA adapters in LLM4WM (§III.C Step 2)—contributes physical knowledge when generating path loss, scatterers, and channel data from RGB/LiDAR input. That premise is asserted in §II.A.3 ('General knowledge transfer') but never isolated experimentally. The comparisons in Figs. 4, 5, 7 and Tables II, III pit GPT-2-based schemes against task-specific baselines with different architectures and parameter counts, so observed gains could come from transformer capacity, MoE-LoRA, or the training protocol rather than from language pretraining. No error bars, seeds, or ablations are reported, so 'significant superiority' is not statistically established. The paper's own §V.B.3 concedes that CSI-only wireless FMs lack native multi-modal capability, but no analogous admission or test addresses the language-to-wireless transfer question. If that transfer does not hold generally, the LLM roadmap loses its foundation and only the from-scratch wireless foundation model roadmap remains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a systematic framework for applying foundation models (FMs) to Synesthesia of Machines (SoM), an AI-native paradigm integrating multi-modal sensing and communication. It categorizes FMs into general-purpose LLMs and domain-specific wireless foundation models, and proposes two design roadmaps. For each roadmap, it offers a step-by-step framework and case studies: LLM4PG and LLM4SG for mechanism exploration, LLM4WM for transceiver design, and WiCo, WiFo, and WiPo for mechanism exploration, transceiver design, and cooperative perception respectively. Preliminary simulations are presented to support the claim that FM-based schemes significantly outperform task-specific models.","tokens_in":29539,"tokens_out":3687,"duration_ms":45546,"significance":"If validated, the proposed categorization and roadmaps would provide a useful design blueprint for 6G research, organizing a fragmented body of work and highlighting where LLMs versus wireless foundation models are more appropriate. The paper makes a strength of being the first to systematically organize FM-based SoM design and includes several concrete case studies with external baselines. However, the empirical evidence is preliminary: figures show single-run comparisons, some with only qualitative image outputs, and no statistical uncertainty is reported. The central premise of Roadmap 1—that language-pretrained GPT-2 provides useful physical knowledge for wireless tasks—is asserted but not experimentally isolated. The framework itself is plausible and potentially significant, but the current evidence base is insufficient to support the strong 'significant superiority' claim in the abstract.","major_comments":[{"comment":"The claim that LLM-based schemes (LLM4PG, LLM4SG, LLM4WM) benefit from 'general knowledge transfer' of language pretraining is never isolated experimentally. The comparisons pit GPT-2 with only LayerNorm parameters (or MoE-LoRA adapters) trainable against task-specific GAN, ResNet, or Cross-Stitch baselines that differ in architecture, parameter count, and training protocol. Since the GPT-2 backbone is frozen, an observed gain could come from transformer capacity, the adaptation modules, or the training procedure rather than from language pretraining. A randomized-initialization ablation—same architecture, same tuning strategy, but with randomly initialized GPT-2 weights—is needed to attribute any improvement to the pretrained representations. Without it, Roadmap 1's core premise remains unverified.","section":"§III.B, §III.C, Figs. 4 and 5"},{"comment":"All reported comparisons are single-run results with no error bars, confidence intervals, or significance tests. Phrases such as 'achieves over 11% higher generalization accuracy' (Fig. 4) and 'significantly outperforms' (Fig. 9) are not statistically established. The paper does not report the number of random seeds, the variance across runs, or the exact fine-tuning dataset sizes and splits. To support the abstract's claim of 'significant superiority of FMs over task-specific models', the authors should provide means and standard deviations over multiple seeds, or equivalent statistical evidence, and specify training hyperparameters and dataset partitions for every case study.","section":"Figs. 4, 5, 7, 8, 9, 10 and Tables II–VI"},{"comment":"WiCo is presented as a 'wireless channel foundation model' but, as described, it is trained on a single dataset (SynthSoM) for a single generation task (path loss or multipath generation). This does not meet the paper's own definition in §II.B.1, which requires pre-training on broad wireless and multi-modal sensing data and adaptability to a wide range of tasks through few-shot or zero-shot learning. The pre-training strategy is described only as 'data augmentation and noise suppression' (Step 3), which is not a specific self-supervised objective. The authors should clarify what distinguishes WiCo from a conventional task-specific generative model, or revise the terminology.","section":"§IV.B, Case Study 1 for Roadmap 2 (WiCo)"},{"comment":"The WiFo comparison in Fig. 9 uses, as a task-specific baseline, a randomly initialized WiFo encoder with otherwise identical components; this is a good controlled ablation for the value of WiFo's pre-training. However, no analogous controlled ablation is provided for the LLM-based schemes, so the WiFo result does not compensate for the missing LLM ablation. Additionally, the WiFo fine-tuning uses 1,500 aligned CSI-RGB pairs but the training/validation split and the number of runs are not given, and the LLM4WM training details (e.g., MoE-LoRA rank, learning rate, number of epochs) are absent. These missing specifications make it impossible to reproduce or independently verify the reported gains.","section":"§IV.C, WiFo case study and §III.C, LLM4WM case study"}],"minor_comments":[{"comment":"The phrase 'self-supervised pertaining' should be 'self-supervised pre-training'.","section":"§II.B.2"},{"comment":"The paper states that GPT-2 is used but does not specify which variant (e.g., GPT-2 small, medium) or how the reported total parameter counts (e.g., 275.70M for LLM4PG) arise. Clarifying the backbone variant and the exact trainable parameter set would help reproducibility.","section":"§III.B, Step 2 and Table II"},{"comment":"The term 'adaption' is used in several places (e.g., §IV.B Step 4 and §IV.B 'adoption') where 'adaptation' is intended; please correct these typos.","section":"§IV.B, Step 4 and elsewhere"},{"comment":"Several figures lack axis labels or legends (e.g., Figs. 4(d) and 4(e) have no labeled axes, and Figs. 9 and 10 do not define NMSE and PSNR axes fully). Adding clear captions with evaluation protocols would improve readability.","section":"Fig. 4, Fig. 9, Fig. 10"},{"comment":"The sentence 'as demonstrated in case study 1 of Section IV-C' appears to be a cross-reference error; the WiFo case study with vision enhancement is described in §IV.C, which is actually 'Case Study 2' of Roadmap 2. Please correct the reference.","section":"§V.B.3"},{"comment":"The text repeatedly writes 'UA V images' (with an extra space) instead of 'UAV images'; please fix this typo.","section":"§IV.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a hybrid survey and case-study paper. The conceptual framework is a useful contribution, but the empirical claims in the abstract and case studies currently exceed the evidence. The most impactful request is a randomized-initialization ablation for the LLM-based schemes and multi-seed reporting with confidence intervals. If the authors prefer to keep the paper as a design framework rather than an empirical benchmarking paper, they could instead soften the 'significant superiority' claim and label the case studies as illustrative. I would also suggest the editor ask the authors to clarify how WiCo and WiPo satisfy the paper's own definition of 'foundation model'."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a useful organizing paper, not a rigorous empirical one. The taxonomy (LLMs vs wireless foundation models) and the two roadmaps are the real contribution; the case studies are mostly the authors' prior work re-presented as illustrations, and the quantitative support is thinner than the abstract suggests.\n\nWhat's actually new: the systematic categorization, the Table VII paradigm comparison, and the step-by-step frameworks (LLM selection plus adaptation; dataset/architecture/pretraining/adaptation for wireless FMs). That is a genuine service to the ISAC/SoM community. WiCo and WiPo appear to be new proposals, and the parameter-count tables are a nice touch.\n\nThe soft spots are in proportion. First, no error bars, seeds, or significance tests appear in any of the figures; 'significant superiority' is asserted from single runs. Second, the LLM roadmap rests on the premise that GPT-2's language pretraining transfers physical structure to path loss, scatterers, and channels. The paper never isolates that premise. Comparing GPT-2 with LayerNorm tuning against a GAN or ResNet confounds architecture, capacity, and training protocol; a randomized-initialization ablation would settle it. The paper itself concedes in Section V.B.3 that CSI-only wireless FMs lack native multi-modal capability, but it never gives the analogous caveat for the language-to-wireless transfer. That is a fair, load-bearing criticism, not a manufactured one. Third, significant chunks are re-presentations of LLM4CP, LLM4SG, LLM4WM, and WiFo; the paper cites them but doesn't clearly mark what is new versus prior.\n\nThe central conceptual argument holds up: foundation models are a reasonable answer to the four SoM challenges, and Roadmap 2 does not depend on the shaky transfer premise. The citation pattern is heavy on self-citations, but mostly to prior work that actually introduced the case studies, so I do not read that as a flaw.\n\nWho gets value: readers new to FM+ISAC who want a map of the landscape and a design pipeline. It deserves a serious referee, but should be sent back for major revision: reframe as a survey/vision paper with the case studies as preliminary illustrations, add the ablation and error bars, release code/data, and tone down 'significant superiority.' I would read the revised version.","headline":"A useful taxonomy and design roadmap for foundation-model-empowered SoM, but the load-bearing claim that LLM pretraining transfers to wireless tasks is asserted, not tested.","tokens_in":30151,"tokens_out":2683,"would_cite":true,"duration_ms":32597,"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":"Foundation models, pre-trained at scale and adaptable to many tasks, are the proposed backbone for 6G's sensing-communication paradigm SoM, giving two concrete design roadmaps that beat task-specific AI models in the reported case studies.","keywords":["Synesthesia of Machines","foundation models","large language models","wireless foundation models","integrated sensing and communications","channel prediction","multi-modal sensing","6G networks"],"falsifier":"Fine-tune GPT-2 on the same SynthSoM tasks after randomly re-initializing its pretrained weights; if the randomized model matches the pretrained model's path-loss and scatterer accuracy, the claimed benefit of language pretraining is absent. Alternatively, evaluate LLM4SG on a held-out city or frequency band never seen in pretraining or fine-tuning; a sharp accuracy drop would contradict the claimed generalization.","tokens_in":29073,"feed_emoji":"📡","tokens_out":4076,"duration_ms":44232,"temperature":0.7,"pith_summary":"The paper argues that foundation models—models pre-trained at scale and adapted to many tasks—are the right backbone for Synesthesia of Machines (SoM), the proposed 6G paradigm that couples multi-modal sensing with wireless communication. Existing task-specific AI models for SoM fail on four counts: scarce aligned datasets, limited modeling capacity, poor generalization, and no task universality. The paper's central claim is that two families of foundation models solve all four: general-purpose large language models (LLMs), adapted by fine-tuning or prompting, and domain-specific wireless foundation models, pre-trained on channel and sensing data. It proposes concrete roadmaps and reports that LLM-based path-loss and scatterer generation, multi-task transceivers, and wireless foundation models (WiCo, WiFo, WiPo) outperform task-specific baselines. If right, SoM system design becomes a choice of backbone plus an adaptation strategy rather than a bespoke network per task.","feed_headline":"Foundation models beat task-specific AI for 6G sensing-communication","feed_subtitle":"Two roadmaps—LLM adaptation or wireless foundation models—cover sensing, transmission, and perception.","key_machinery":"The load-bearing mechanism is pre-training plus task adaptation. For LLMs, frozen GPT-2 weights act as a general-knowledge prior; only LayerNorm parameters (LLM4PG/LLM4SG) or LoRA experts (LLM4WM) are trained, converting camera, LiDAR, or CSI tokens into channel outputs. For wireless foundation models, self-supervised objectives—masked reconstruction for WiFo, cross-modal generation for WiCo, heterogeneous token reconstruction for WiPo—produce a shared backbone that is then frozen or lightly fine-tuned. The SynthSoM dataset supplies the aligned multi-modal sensing-communication pairs that both roadmaps train and evaluate on.","core_discovery":"On its own terms, the paper establishes a systematic framework: foundation models for SoM fall into two classes—general-purpose LLMs and SoM-specific wireless foundation models—and each class addresses the four SoM challenges through a distinct mechanism. For LLMs, the mechanism is transfer: pre-trained language representations, after light fine-tuning (LayerNorm tuning for generation tasks, MoE-LoRA for multi-task transceivers), generate path loss maps, scatterers, and channel predictions from camera and LiDAR inputs better than GAN or ResNet baselines. For wireless foundation models, the mechanism is self-supervised pre-training on channel or heterogeneous multi-modal data, followed by zero-shot or few-shot adaptation; WiCo generates path loss and multipath parameters from images, WiFo predicts channels from CSI with or without RGB, and WiPo transmits modality-agnostic features for cooperative perception. Each case study reports that the foundation-model scheme beats the task-specific model on accuracy or generalization with comparable inference time and parameter count. The paper positions this as the first systematic categorization and design pipeline for FM-empowered SoM.","pith_inferences":["The reported gains are preliminary and benchmarked on SynthSoM; a stronger test would randomize the pretrained LLM weights to see how much of the advantage comes from language pretraining versus the fine-tuning procedure itself.","If the transfer premise holds broadly, the same roadmap could apply to other radio tasks such as spectrum sensing and beam management, where labeled data is scarce but visual or LiDAR context is available.","The wireless-foundation-model roadmap may eventually make LLM-based SoM unnecessary for physical-layer tasks, since WiFo and WiCo match or exceed LLM results with fewer parameters."],"forward_implications":["If FM transfer holds, scarce aligned sensing-communication datasets stop being the bottleneck: few-shot adaptation replaces large task-specific collections.","A single LLM or wireless foundation model can serve multiple SoM tasks, cutting deployed model count and storage overhead.","Zero-shot channel prediction becomes feasible in new scenarios, removing the need for per-scenario retraining.","SoM mechanism exploration can be automated: LLMs and WiCo generate path loss and scatterers from visual data, supporting transceiver design without full channel measurement."],"supporting_citations":[{"why":"Supplies the GPT-2 backbone whose pretrained representations carry the general-knowledge prior for LLM4PG, LLM4SG, and LLM4WM.","marker":"[121]"},{"why":"Provides the earlier LLM4CP evidence that fine-tuned LLMs outperform task-specific and model-based channel prediction with few CSI samples.","marker":"[66]"},{"why":"Supplies the MoE-LoRA multi-task fine-tuning method used in the LLM4WM transceiver case study.","marker":"[72]"},{"why":"Defines WiFo, the wireless foundation model for channel prediction that anchors the second roadmap's transceiver case study and its zero-shot results.","marker":"[104]"},{"why":"Provides the SynthSoM dataset of aligned RGB, LiDAR, radar, and channel data used to train and evaluate the proposed schemes.","marker":"[8]"},{"why":"Introduces LLM-based scatterer generation, the SoM mechanism exploration extended as LLM4SG.","marker":"[120]"},{"why":"Defines Synesthesia of Machines, the paradigm whose four challenges the paper claims foundation models solve.","marker":"[4]"},{"why":"Establishes the foundation-model concept of pre-training plus task adaptation on which both roadmaps rest.","marker":"[11]"}],"fun_headline_variants":["6G SoM: Foundation models beat task-specific AI via two roadmaps","LLM and wireless foundation models outperform in 6G SoM","First systematic FM design for 6G SoM beats task-specific AI","Two FM roadmaps—LLM and wireless—excel in 6G sensing-communication"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that what a language model learned from text and images helps it understand radio signals, so that lightly fine-tuning GPT-2 produces accurate path loss, scatterers, and channel predictions from camera and LiDAR data.","fun_headline_variants_meta":{"raw":{"variants":["6G SoM: Foundation models beat task-specific AI via two roadmaps","LLM and wireless foundation models outperform in 6G SoM","First systematic FM design for 6G SoM beats task-specific AI","Two FM roadmaps—LLM and wireless—excel in 6G sensing-communication"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000455,"raw_usage":{"total_tokens":2362,"prompt_tokens":1101,"completion_tokens":1261,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":717,"completion_tokens_details":{"reasoning_tokens":1189}},"tokens_in":717,"tokens_out":1261,"duration_ms":13187,"temperature":1.0,"reasoning_tokens":1189,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:28:36.267774+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fine-tune GPT-2 on the same SynthSoM tasks after randomly re-initializing its pretrained weights; if the randomized model matches the pretrained model's path-loss and scatterer accuracy, the claimed benefit of language pretraining is absent. Alternatively, evaluate LLM4SG on a held-out city or frequency band never seen in pretraining or fine-tuning; a sharp accuracy drop would contradict the claimed generalization.","supporting_citations":[],"review_version":1}