{"id":"6fc2152a-a3f5-49d1-a588-9d98773aaae5","arxiv_id":"2412.09041","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A position paper proposing that wireless intelligence should be built natively from radio physics, not transferred from large language models.","lead":"This position paper argues that wireless networks need their own kind of big AI model, built around radio signals and physical laws, rather than simply borrowing large language models. It reviews current work and lays out a roadmap for a 'wireless native' model that could handle many tasks at once.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inference in Sec. 3.2 that cross-task and cross-scenario generalization implies internal capture of electromagnetic laws is unsupported and is more foundational than the scaling assumption.","rationale":"The reader's verdict is CONDITIONAL and identifies the scaling assumption in Sec. 5.3 as the weakest point. I agree that scaling and emergence are asserted without quantitative support, but I see a more fundamental weakness located earlier in the argument: the wBAIM definition and evaluation criteria rest on the premise that behavioral generalization across tasks and scenarios certifies acquisition of electromagnetic laws (Sec. 3.2). This is an inverse inference with known failure modes such as multiple realizability and shortcut learning. The scaling assumption only becomes relevant if the model must actually scale to encode the laws; if the laws are as simple as Sec. 4.2 claims, a physics-driven model might generalize without large scale. The proposed linear-probing experiment would test the content of the shared representation rather than only its predictive accuracy, directly targeting the paper's central equivalence. Since the manuscript is explicitly a position paper, this unsupported but testable premise warrants a conditional rather than a reject verdict, so the reader's verdict remains appropriate.","tokens_in":11655,"tokens_out":7609,"duration_ms":81991,"concrete_test":"Test the Sec. 3.2 equivalence directly on the existing CMixer multi-task, multi-scenario setup from [19]. Train two models of equal capacity: (i) the reported physics-inspired CMixer, and (ii) a matched-capacity Transformer or MLP baseline with no physics-inspired inductive bias, on the same tasks and scenarios. Evaluate both on held-out scenarios. Then apply linear probing to the shared representations of both models to measure how accurately they decode physical parameters such as multipath delays, angle-of-arrival, and Doppler spread. If the baseline matches or exceeds CMixer's cross-scenario generalization while its representations do not decode physical parameters, then generalization does not imply electromagnetic-law capture and the Sec. 3.2 inference fails. If instead all generalizing models necessarily decode physical parameters, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is not merely that wBAIM will scale but that a single model generalizing across multiple wireless tasks and scenarios 'must inherently capture the commonalities among them, that is, the general electromagnetic laws' (Sec. 3.2). This inference is load-bearing because it justifies the wBAIM definition and the implicit evaluation via multi-task and multi-scenario performance. The inference is not logically secure: multi-task or multi-scenario generalization can arise from shared but non-physical feature statistics, task-specific decoders over a domain-agnostic representation, or dataset biases such as scenario IDs and common measurement artifacts, without the model internalizing Maxwell-based laws. The paper provides no experimental or analytical evidence that the commonalities learned by wireless models are physical laws rather than statistical shortcuts. Furthermore, Sec. 4.2 states that the fundamental laws of electromagnetic waves are 'relatively simple,' which makes it less obvious that scale is what forces law-like representations. Since this equivalence is asserted rather than demonstrated, the roadmap's foundation is weaker than the paper acknowledges. The scaling concern in Sec. 5.3 is downstream: even if scaling works, it would not establish that the resulting model has learned electromagnetic laws.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This position paper argues that applying large language model (LLM) technology to wireless systems is insufficient and that a wireless-native big AI model (wBAIM) should be developed instead. It reviews 36 publications on BAIM for wireless, categorizes them into wireless LLM, LLM-based wireless agents, and wBAIM, and proposes that wireless intelligence is fundamentally different from language intelligence because it is hyper-cognitive rather than human-centric. The paper then identifies peculiarities of wBAIM in data attributes, model functionality, and applied scenarios, and proposes five methodologies: hybrid data collection, physics-driven learning, wireless scaling laws, structural prompting, and collaboration with AI agents. The central thesis is that wBAIM should learn electromagnetic laws directly from wireless data and will generalize across tasks and scenarios as a result.","tokens_in":11946,"tokens_out":2154,"duration_ms":23016,"significance":"If the central thesis holds, this paper would provide a useful conceptual foundation for a shift from task-specific wireless AI models to unified foundation models for 6G and electromagnetic sensing. The paper's strengths are its clear research questions, systematic taxonomy of current BAIM-for-wireless paradigms, and explicit identification of differences between language and wireless intelligence. It also gives specific methodological directions, such as structural prompts and physics-driven learning, that are actionable. The paper is honest about the early stage of wBAIM research and does not overclaim experimental results. Its main weakness is that the logical link between cross-task generalization and internalization of electromagnetic laws is asserted rather than demonstrated, and the scaling-law argument rests on analogy to LLMs without wireless-specific evidence.","major_comments":[{"comment":"The statement that 'If a single model can be used to generalize across multiple wireless tasks and scenarios, it must inherently capture the commonalities among them—that is, the general electromagnetic laws' is an unsupported logical inference. Generalization across tasks and scenarios can arise from shared but non-physical feature statistics, task-specific decoders over a domain-agnostic representation, or dataset biases such as scenario identifiers and common measurement artifacts. The paper provides no experimental or analytical evidence that the commonalities learned by wireless models are physical laws. Since this equivalence is used to justify the wBAIM definition and its implicit evaluation metrics, it is load-bearing and needs to be either supported with evidence (for example, probing experiments or out-of-distribution tests that distinguish physics-consistent predictions from statistical shortcuts) or reformulated as a hypothesis rather than a necessity.","section":"Section 3.2"},{"comment":"The claim that 'increasing model size is also essential for the performance of wireless models, especially in terms of generalization' and that jointly processing multi-modal, multi-user, multi-scenario data will generate the 'more is different' emergence effect is asserted without wireless-specific evidence. The paper cites reference [19] as preliminary confirmation, but no scaling law, fitted exponent, or experimental demonstration is provided. For a position paper it is acceptable to propose scaling as a hypothesis, but the current wording states it as a principle and therefore overstates the support. This is load-bearing because the entire methodology of building a large wBAIM depends on this transfer of scaling behavior from LLMs to wireless models.","section":"Section 5.3"},{"comment":"There is a tension between the claim in Section 4.2 that 'the fundamental laws of electromagnetic waves are relatively simple' and the claim in Section 3.2 that a single model generalizing across tasks and scenarios must capture these laws. If the laws are simple, it is not obvious why very large models and massive pre-training are necessary to learn them; compact physics-informed models might suffice. The paper should reconcile these statements by explaining why, despite the simplicity of the underlying laws, scale and data diversity are still needed—for example, because the complexity lies in the coupling with scattering environments and transmission mechanisms. As written, the two sections pull in opposite directions and weaken the scaling argument.","section":"Section 4.2"}],"minor_comments":[{"comment":"The statistical text under Figure 3 contains a clear typesetting artifact ('/uni00000033/uni00000055/...') that renders as garbled code instead of readable labels; this needs to be fixed in the production version.","section":"Figure 3"},{"comment":"The title page says 'POSITION P APER' with an extra space; also, the corresponding author email is written as 'ning ming@zju.edu.cn', which appears to be a spacing error.","section":"Title page"},{"comment":"The acronym 'BAIM' is used both as a general term for big AI models and as part of 'wBAIM'; the paper would benefit from a brief definition of BAIM in the abstract or first use, since the abstract introduces it without explanation.","section":"Abstract and Introduction"},{"comment":"The phrase 'L VMs' contains an unintended space; it should read 'LVMs'.","section":"Section 4.1"},{"comment":"Reference [34] contains the typo 'telecom-specfic'; please correct to 'telecom-specific'.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior work [19], which defines the three key wBAIM features and is then used as the basis for arguing wBAIM's necessity. This is not inherently improper for a position paper, but the editor should be aware that the central conceptual framework is self-referential. The paper's fit with the journal seems appropriate for a position paper, though the technical depth is modest. The major revision should focus on substantiating or softening the inference in Section 3.2 and the scaling claims in Section 5.3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a position paper arguing that wireless big AI models should be native to electromagnetic physics, not borrowed from LLMs. The genuinely new idea is the human-centric vs hyper-cognitive intelligence distinction: language models learn from human-generated data and feedback, while wireless models should learn from physical phenomena and system feedback. That framing is useful and goes beyond the usual 'we need foundation models for wireless' pitch. The paper also maps four LLM success factors (data, learning paradigm, scaling, prompting) to wireless-specific counterparts, which gives the roadmap structure. The literature review is current and reasonably thorough for a fast-moving area.\n\nWhere it is soft: the load-bearing inference in Sec. 3.2. The paper claims that if one model generalizes across wireless tasks and scenarios, it 'must inherently capture' the general electromagnetic laws. That is not established. A model could generalize via shared statistical structure, task-specific decoders, or dataset biases without internalizing Maxwell-like laws. The stress-test note is right that this is more foundational than the scaling-law assumption. The paper treats 'compression is intelligence' as self-evident, but it is a hypothesis. Similarly, Sec. 5.3 asserts that increasing model size is essential, citing [19] as 'preliminary confirmation' - there is no scaling law or experiment here. These are appropriate as research directions for a position paper, but they are presented with more certainty than the evidence supports.\n\nAlso, the three wBAIM features (multi-task, multi-scenario, all-in-one scheduling) come from the authors' own prior work [19], and the paper then uses those as the basis for necessity. That is a minor circularity for a roadmap paper, not a fatal one.\n\nWho is this for: researchers in wireless AI who want a structured argument for why wireless foundation models need physical grounding, and a good entry-point survey. It deserves a serious referee: the framing is valuable even if the evidence is mostly illustrative. I would engage with it. For peer review, I would recommend accepting with revisions that temper the 'must' language, add a discussion of alternative generalization mechanisms, and maybe include a preliminary scaling study or at least a more careful formulation of what a 'wireless scaling law' would mean.","headline":"Useful roadmap with a novel human-centric vs hyper-cognitive framing, but the core inference that multi-task generalization implies internalizing EM laws is asserted rather than shown.","tokens_in":12388,"tokens_out":1908,"would_cite":true,"duration_ms":18683,"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":"The paper argues that general wireless AI requires a wireless-native big model that learns electromagnetic laws directly, rather than adapting large language models.","keywords":["wireless native AI","big AI model","foundation model","6G","electromagnetic laws","hyper-cognitive intelligence","wireless scaling laws","physics-driven learning"],"falsifier":"Run a controlled scaling study on real measured wireless channel data: train a family of wBAIMs with increasing parameter counts and increasingly diverse multi-scenario training sets, and check whether performance on an unseen scenario improves smoothly or abruptly. If no such improvement appears, or if gains saturate, the paper's wireless scaling-law premise is falsified.","tokens_in":11484,"feed_emoji":"📡","tokens_out":3832,"duration_ms":37554,"temperature":0.7,"pith_summary":"The paper argues that the route to general wireless AI is not to adapt large language models, but to build a wireless-native big AI model (wBAIM) that learns from electromagnetic phenomena directly. It claims language intelligence and wireless intelligence are different in kind: language is human-centric, while electromagnetic systems are hyper-cognitive, meaning they exceed direct human understanding. Because wireless tasks and scenarios share underlying electromagnetic laws, a single model that generalizes across tasks must be compressing those laws. The paper surveys current paradigms, identifies the peculiarities of wireless data, models, and applications, and proposes four methodological pillars: hybrid data collection, a physics-driven learning paradigm, wireless scaling laws, and structural prompting. If the vision is right, 6G wireless intelligence would shift from task-specific models to a unified foundation model that also supports radar and remote sensing.","feed_headline":"Build wireless AI on electromagnetic laws, not language","feed_subtitle":"A position paper argues a wireless-native big AI model can unify tasks and scenarios by capturing general electromagnetic laws.","key_machinery":"The load-bearing distinction is between human-centric and hyper-cognitive intelligence, coupled with the identity that cross-task, cross-scenario generalization implies learning common electromagnetic laws. The paper's proposed machinery for realizing this is a synthesis of four LLM evolutionary drivers mapped to wireless: hybrid data collection that mixes multi-scenario measurements with simulation; a physics-driven learning paradigm that injects known electromagnetic structure into model design; wireless scaling laws in which jointly processing multimodal, multi-user, multi-scenario data is what produces the 'more is different' emergence effect; and structural prompting, where prompts are strictly structured multimodal data rather than free-form language, so the model can adapt without retraining.","core_discovery":"The central claim is that a wireless-native big AI model should be developed as a hyper-cognitive model whose training signal is observed electromagnetic phenomena rather than human-generated text. On this view, the reason LLMs cannot be transplanted to wireless is not missing domain knowledge but a mismatch in intelligence orientation: language models imitate human cognition, whereas electromagnetic systems require intelligence that surpasses it. The paper asserts that if a single model generalizes across multiple wireless tasks and scenarios, it must inherently capture the commonalities among them, namely the general electromagnetic laws, which it calls 'compression is intelligence' in the wireless context. From this it follows that wBAIM's defining features of multi-task integration, multi-scenario unification, and all-in-one scheduling are not conveniences but implicit evidence that the model has learned real physics.","pith_inferences":["If the compression-is-intelligence claim holds for electromagnetics, then wBAIM performance on a held-out task becomes a measurable proxy for whether the model has internalized physical laws; one could test this by probing latent representations against analytic channel models.","The paper's wireless scaling-law argument could be made testable by measuring whether multi-scenario pre-training produces error drops on new scenarios that fit a power law in model size and data mixing ratio, but the paper does not provide such exponents.","The hyper-cognitive framing suggests that wBAIM, if successful, would be a template for other AI-for-science domains such as weather, materials, or fluid dynamics, where the teacher is the system itself rather than human labels.","The decentralized constraint implies that wBAIM pre-training may need to proceed through model-interaction schemes like federated learning, which the paper mentions only briefly."],"forward_implications":["LLM-based wireless approaches, including wireless LLMs and LLM-based agents, will remain limited to interaction and simple tasks; hard wireless problems such as high-rate transmission and scenario sensing will need a native model.","Training wBAIM will require mixing measured and simulated data, with simulation providing high-fidelity multipath components and unbiased user locations that are difficult to obtain from real measurements.","Adaptation of wBAIM will rely on structural prompts rather than fine-tuning, because retraining is too slow and resource-hungry for wireless nodes with frequently changing environments.","A successful wBAIM should be usable beyond communications, providing technical support to remote sensing and radar systems.","Collaborating wBAIM with AI-agent modules for planning, observation, memory, and small tool libraries would turn it into an active intelligent wireless brain rather than a passive predictor."],"supporting_citations":[{"why":"Defines wBAIM's three key indicators (multi-task integration, multi-scenario unification, all-in-one scheduling) and provides the CMixer case study the paper builds on.","marker":"[19]"},{"why":"Provides the foundation-model framing and evidence that pre-trained models generalize across language tasks, the benchmark the paper wants to replicate for wireless.","marker":"[15]"},{"why":"Illustrates the physics-inspired approach to wireless model design that wBAIM's learning paradigm extends.","marker":"[8]"},{"why":"Channel deduction example shows how reorganizing modules around physical correlations enables low-overhead channel acquisition.","marker":"[13]"},{"why":"Supplies the data-distribution similarity and diversity assessment used to guide multi-scenario data mixing.","marker":"[57]"},{"why":"Shows wireless data treated as sequences rather than images, an example of tailoring representation learning to wireless structure.","marker":"[7]"}],"fun_headline_variants":["Wireless AI needs electromagnetism, not language","Train wireless AI on EM laws, not text","Wireless-native AI: physics over language","Forget LLMs: wireless AI learns from EM fields"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that big-model scaling and joint multi-modal, multi-user, multi-scenario training will generate the same 'more is different' emergence in wireless models that it did in language models; the paper presents this as a principle with only preliminary confirmation and no scaling-law fit.","fun_headline_variants_meta":{"raw":{"variants":["Wireless AI needs electromagnetism, not language","Train wireless AI on EM laws, not text","Wireless-native AI: physics over language","Forget LLMs: wireless AI learns from EM fields"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1337,"prompt_tokens":857,"completion_tokens":480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":420}},"tokens_in":473,"tokens_out":480,"duration_ms":5140,"temperature":1.0,"reasoning_tokens":420,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:21:10.389211+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled scaling study on real measured wireless channel data: train a family of wBAIMs with increasing parameter counts and increasingly diverse multi-scenario training sets, and check whether performance on an unseen scenario improves smoothly or abruptly. If no such improvement appears, or if gains saturate, the paper's wireless scaling-law premise is falsified.","supporting_citations":[{"cited_title":"Big AI models for 6G wireless networks: Opportunities, challenges, and research directions","cited_arxiv_id":null,"evidence_quote":"Defines wBAIM's three key indicators (multi-task integration, multi-scenario unification, all-in-one scheduling) and provides the CMixer case study the paper builds on."},{"cited_title":"C-GRBFnet: A physics-inspired generative deep neural network for channel representation and prediction","cited_arxiv_id":null,"evidence_quote":"Illustrates the physics-inspired approach to wireless model design that wBAIM's learning paradigm extends."},{"cited_title":"Channel Deduction: A New Learning Framework to Acquire Channel from Outdated Samples and Coarse Estimate","cited_arxiv_id":"2403.19409","evidence_quote":"Channel deduction example shows how reorganizing modules around physical correlations enables low-overhead channel acquisition."},{"cited_title":"Assessing air-interface dataset similarity and diversity for AI-enabled wireless communications","cited_arxiv_id":null,"evidence_quote":"Supplies the data-distribution similarity and diversity assessment used to guide multi-scenario data mixing."},{"cited_title":"Viewing channel as sequence rather than image: A 2-D Seq2Seq approach for efficient MIMO-OFDM CSI feedback","cited_arxiv_id":null,"evidence_quote":"Shows wireless data treated as sequences rather than images, an example of tailoring representation learning to wireless structure."}],"review_version":1}