{"id":"7f35eae7-2609-4d8f-93ce-0a87a34cd9b5","arxiv_id":"2608.00255","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A cross-architecture survey that organizes AI methods for movable, fluid, and pinching antennas by the joint optimization problem they solve.","lead":"This paper surveys recent work applying artificial intelligence to three types of reconfigurable antennas: movable, fluid, and pinching antenna systems. It is a useful map of the 6G literature for researchers choosing learning methods, but it introduces no new measurements, algorithms, or derivations.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim that AI makes spatial reconfigurability practical rests on simulation-only evidence; the survey's own §VI.A.4 admits no measured datasets, leaving the 'practical' predicate unsupported.","rationale":"The reader's weakest assumption is that offline-trained AI policies can be made robust enough for real-time reconfiguration in practice, and the reader already reaches CONDITIONAL based partly on the sim-to-real gap. This stress-test independently locates the same concern in the survey's own self-identified limitation in §VI.A.4 and in the repeated lessons-section caveats. The concrete check is a systematic audit of the surveyed evidence base rather than a new experiment, because the central claim is about what the literature supports. The concern does not invalidate the survey's taxonomic contribution — the unified organization of MA/FAS/PASS under AI-oriented categories is genuinely useful — but it does show that the 'practical' component of the headline claim is aspirational, not established. Since the reader's verdict is already CONDITIONAL and this concern reinforces rather than redirects that verdict, no change is required.","tokens_in":46982,"tokens_out":4234,"duration_ms":46172,"concrete_test":"Audit Tables IV–VIII systematically. For each cited work, record (i) evaluation data source: analytic channel simulation, standardized ray-tracing, measured dataset, or hardware testbed; (ii) CSI assumption: perfect, partial/imperfect, or estimation-free; (iii) whether online inference latency/energy of the AI controller is reported. Then recompute the survey's synthesis restricted to works with measured/hardware validation and hardware-aware imperfect-CSI evaluation. If the restricted set is empty — as §VI.A.4 suggests — the 'practical' claim should be explicitly downgraded to 'simulation-oriented design framework,' and the conditional verdict remains pending hardware benchmarks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim in §I is that AI is 'not merely an additional technique, but a common design framework for making spatial reconfigurability practical.' For this to hold, the surveyed learning policies must work under real CSI acquisition overhead, actuation limits, and sim-to-real channel mismatch. The survey does not establish this. It explicitly flags the gap in §VI.A.4: 'the limited availability of measured MA, FAS, and PASS datasets means that most policies are trained and evaluated using simulation models,' and §VI.A.4 further notes that large AI backbones may consume the latency/memory/energy budgets that spatial reconfiguration is supposed to reduce. The lessons sections (§III.F, §IV.F, §V.D) repeat the same caveat: simplified channel models, perfect CSI assumptions, fixed system dimensions, and offline training distributions. The physical channel models used throughout the reviewed works — e.g., the field-response model in Eqs. (3)–(8), Clarke's isotropic scattering in Eq. (15), and the lossless-waveguide PASS model in Eq. (19) — omit actuator delay, switching cost, mutual coupling beyond a minimum-spacing constraint, and inference energy. Those are exactly the factors named in §VI.A.4 as unresolved. A survey may legitimately organize simulation research, but the word 'practical' in the central claim goes beyond taxonomy: it asserts deployability. The evidence base in Tables IV–VIII contains no hardware-validated counterexample, so the claim overstates what is currently known.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper surveys AI-enabled design for three spatially reconfigurable antenna paradigms—movable antennas, fluid antenna systems, and pinching antenna systems—with a unified taxonomy organized by joint optimization problem (beamforming, channel estimation, ISAC, security, resource allocation, etc.). It summarizes the operating principles and standard channel models for each architecture, reviews a large body of recent learning-based works, compares the roles of DL/DRL/GNN/Transformer/LLM/KKT-guided methods, and closes with open challenges and future directions. The central claim, stated in Section I, is that AI is 'not merely an additional technique, but a common design framework for making spatial reconfigurability practical.'","tokens_in":47375,"tokens_out":3087,"duration_ms":34333,"significance":"The survey fills a genuine gap: prior surveys treat MA, FAS, or PASS separately, while this paper provides a side-by-side treatment of all three under a learning-oriented taxonomy. The channel models quoted in Eqs. (3)–(8), (15), and (19) are standard and appear correctly transcribed, and the summary tables (Tables IV–VIII) give a useful map of recent work. The manuscript is candid about many limitations, especially in Section VI.A.4, which flags the absence of measured datasets and the sim-to-real gap. If the result is taken as a survey of simulation-based AI research, it is a useful reference. However, the load-bearing framing that AI makes spatial reconfigurability 'practical' goes beyond what the surveyed evidence supports, because the evidence base is almost entirely simulation-based and the paper's own caveats undercut the deployability predicate.","major_comments":[{"comment":"The central claim that AI is 'a common design framework for making spatial reconfigurability practical' is not supported by the evidence assembled in the survey. Section VI.A.4 states that 'the limited availability of measured MA, FAS, and PASS datasets means that most policies are trained and evaluated using simulation models,' and Sections III.F, IV.F, and V.D repeat caveats about simplified channel models, perfect CSI, fixed dimensions, and offline training distributions. None of the cited works provides a hardware-validated counterexample. The 'practical' predicate should either be qualified to 'practical in simulation-based design' or the paper should add a dedicated assessment of hardware testbeds, measured channels, and deployment evidence. This is load-bearing because the Introduction's strongest claim rests on it.","section":"Section I vs. Section VI.A.4"},{"comment":"The paper argues that AI is 'indispensable' because MA/FAS/PASS problems are non-convex, high-dimensional, and time-varying. But it does not compare learned policies against strong classical solvers under equal computational budgets. The claim in Section II.D.a that SCA and AO are 'thwarted' without strong initialization is contradicted by works cited in the survey itself: [88] provides a robust AO solution for MA-enabled vehicular networks, and [12] presents an optimization-based algorithm with provable convergence to stationary points. Many numerical comparisons in Sections III–V use baselines such as FPA, random selection, or WMMSE with random positions, which do not establish that AI is necessary or superior to a well-initialized classical method. The 'necessity' framing should be softened to 'attractive practical/complexity-motivated alternative' unless the survey provides a systema","section":"Section II.D and Section VI.A.5"},{"comment":"The survey repeatedly reports specific numerical gains—e.g., 'from about 2.8 to about 3.5' in Section III.A, '4.23 bits/s/Hz' in Section III.A, '23.7%, 11.7%, and 9%' in Section V.B—without a common benchmark protocol. These numbers come from heterogeneous papers with different channel models, SNRs, system dimensions, and baseline definitions, so their aggregation is not evidence for a 'common design framework.' If the comparative claim is to be load-bearing, the survey should either state explicit inclusion/exclusion criteria for performance numbers, provide a uniform re-evaluation setup, or clearly mark these as illustrative single-paper results rather than cross-paper comparisons.","section":"Tables IV–VIII and performance numbers in Sections III–V"}],"minor_comments":[{"comment":"The paragraph describing [65] appears twice: the text after the [65] citation in Section III.B is essentially repeated verbatim later in the same subsection. Please remove the duplication.","section":"Section III.B"},{"comment":"Equation (19) contains a typo: 'where where q_{w,m} and u_k denote...' should read 'where q_{w,m} and u_k denote...'.","section":"Section II.C, Eq. (19)"},{"comment":"The manuscript has several spacing and capitalization inconsistencies: 'UA V' should be 'UAV', 'accross' should be 'across', 'Section' is sometimes capitalized and sometimes not, and 'PASSs' is used as a plural in some places. A copyedit pass would improve readability.","section":"Throughout"},{"comment":"The reference list relies heavily on arXiv preprints and very recent '2026' papers. For a survey, it would be helpful to mark which results are peer-reviewed and to note the preprint status more systematically, so readers can gauge maturity.","section":"Section I and References"},{"comment":"The 'Representative work' row in Table III cites only a few references per architecture; consider adding cross-references to the corresponding summary tables (Tables IV–VIII) for traceability.","section":"Table III"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable survey but its central framing overclaims practical deployment. I would ask the authors to either qualify the 'practical' claim explicitly or add a dedicated section on hardware/testbed evidence. The self-citation density is noticeable but not disqualifying; however, the heavy reliance on 2026 arXiv preprints is worth probing during revision to verify that the surveyed results actually exist in the cited form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth your time if you work on any of these antenna architectures. The paper is a survey, not a new result, but it does something genuinely useful: it organizes the AI-for-MA, AI-for-FAS, and AI-for-PASS literatures under one taxonomy keyed to optimization objectives, with comparison tables that make cross-architecture similarities and differences easy to see. The physical-layer channel models quoted are standard and correctly transcribed. The lessons-learned sections in each chapter are honest about common failure modes. The paper also explicitly flags the biggest limitation: most policies are trained and evaluated on simulated channels, with no measured datasets and an acknowledged sim-to-real gap. That admission is more than many surveys offer, and it counts for something.\n\nThe main soft spot is the framing, not the content. The introduction claims AI is a 'common design framework for making spatial reconfigurability practical.' That word 'practical' is not supported by the evidence in the survey, and the stress-test is right to call it out. The paper itself says in Section VI.A.4 that 'the limited availability of measured MA, FAS, and PASS datasets means that most policies are trained and evaluated using simulation models.' A survey can legitimately organize simulation research, but it should say 'promising avenue' rather than 'practical.' This is a one-line fix: the central organizational claim—that AI provides a useful unified lens for these systems—is well supported. The overreach is only in the adjective.\n\nTwo smaller editorial issues. Section III.B contains a duplicated paragraph for [65] and another for [67]; the second copy should be deleted. And Section II.D calls AI 'indispensable' for all three systems, which is stronger than the evidence warrants; the paper later qualifies this, but the earlier claim reads as hype.\n\nWho is this for? A newcomer wanting a map of the field, or a researcher looking for open problems. It deserves a serious referee: a well-done survey of a rapidly growing area with a useful taxonomy is worth reviewing, though the authors should fix the duplication and soften the 'practical' language before publication.","headline":"Competent and useful survey with a cross-architecture AI taxonomy, but the 'practical' framing overstates a simulation-only evidence base that the paper itself admits.","tokens_in":47787,"tokens_out":1786,"would_cite":false,"duration_ms":19342,"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":"Artificial intelligence is the common design framework that can make movable, fluid, and pinching antennas practical.","keywords":["movable antennas","fluid antenna systems","pinching antennas","deep reinforcement learning","beamforming","channel estimation","integrated sensing and communication","6G"],"falsifier":"A controlled hardware testbed would settle it: train any of the surveyed MA, FAS, or PASS policies on the standard simulation channels, then run them on measured channels in the same deployment. If the learned policy's throughput, secrecy rate, or sensing accuracy is no better than a fixed-position baseline once acquisition overhead and actuator/switching latency are included, the common-framework claim fails. The paper itself pinpoints this gap: most policies are trained and evaluated in simulation, so the test is whether the gains survive transfer.","tokens_in":46943,"feed_emoji":"📡","tokens_out":5274,"duration_ms":50656,"temperature":0.7,"pith_summary":"This paper tries to establish that three 6G antenna concepts — movable antennas, fluid antenna systems, and pinching antenna systems — are one problem viewed three ways: how to exploit spatial reconfigurability as an extra degree of freedom. Its unifying claim is that AI is the design framework that makes this reconfigurability practical, since the joint optimization of antenna configuration with beamforming, channel estimation, sensing, security, and resource allocation is high-dimensional, non-convex, mixed discrete-continuous, and time-varying. The survey supports this by classifying recent work by optimization objective and by mapping each antenna architecture's physics to the suitable AI tool: continuous position control for MAs, discrete port selection for FASs, and waveguide-constrained near-field placement for PASSs. A sympathetic reader would care because the promised payoff is millisecond-scale learned decisions instead of repeated optimization, at the price of trusting simulation-trained policies on real hardware.","feed_headline":"Movable, fluid, and pinching antennas share one AI playbook","feed_subtitle":"Survey argues learned policies can replace repeated optimization across beamforming, sensing, security, and resource allocation.","key_machinery":"The mechanism carrying the argument is the spatial configuration variable, defined differently in each architecture: a continuous position vector for MA, a discrete port index for FAS, and a continuous longitudinal coordinate along a dielectric waveguide for PASS. Because the channel depends on this variable, every design task becomes a joint optimization whose difficulty grows with the configuration space. The paper's taxonomy — grouping studies by the coupling problem they solve, and then by the AI model family used — is what allows the authors to treat the three architectures as instances of one framework. Structure-guided learning is the key bridge: KKT-guided Transformers for PASS learn","core_discovery":"The paper's central claim is that AI is not merely an additional technique but a common design framework for making spatial reconfigurability practical. Across the three architectures, antenna configuration is coupled with channel acquisition, beamforming, user scheduling, power control, trajectory design, phase-shift optimization, and task-oriented resource allocation; these problems are high-dimensional, non-convex, mixed discrete-continuous, and time-varying. The survey argues that learned policies — deep learning and Transformers for fast prediction, deep reinforcement learning and multi-agent RL for sequential decisions, graph neural networks for scalability, large language models for c","pith_inferences":["If the common-framework thesis holds, the next natural experiment is a single policy that controls a hybrid MA–FAS–PASS aperture; the paper lists hybrid architectures as a future direction but does not test them.","The reported gains are simulation-based; a measured-channel benchmark could show where the gap to fixed arrays shrinks once channel-estimation overhead, actuator delay, and switching energy are counted.","The energy cost of large AI backbones (LLMs) may cancel the energy savings of spatial reconfiguration; the paper flags this, implying lightweight models and accelerators are prerequisites, not optional.","Standardized control interfaces that expose antenna configuration as a programmable network resource would be the concrete route from this framework to deployment."],"forward_implications":["Learned inference can cut decision latency from hundreds of milliseconds to a few milliseconds (e.g., 7.89 ms vs 644 ms for a movable-antenna positioning/beamforming example), making real-time reconfiguration feasible within a coherence interval.","Sparse-observation extrapolation — reconstructing full-port FAS channel state from 5–10% observed ports — can slash pilot overhead, removing a main barrier to port-domain reconfiguration.","Structure-guided learning (KKT-embedded Transformers, deep-unfolded WMMSE) can beat both black-box learning and classical optimization in sum rate with millisecond inference, as reported for pinching-antenna systems.","A unified spatial-variable formulation of MA positions, FAS port states, and PA coordinates would enable transfer learning across architectures, reducing retraining cost for new deployments.","DRL and MARL can handle dynamic and distributed cases — UAV trajectories, cell-free networks, resilient operation — where repeated optimization is too slow."],"fun_headline_variants":["AI unifies control of movable, fluid, and pinching antennas","One AI playbook for antennas that move, flow, and pinch","Survey: AI learns to reconfigure antennas with speed and adaptability","AI techniques compared across three reconfigurable antenna types","Spatial antenna flexibility meets AI for 6G wireless"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that offline-trained AI policies, developed largely on simulated channel models, can be robust and fast enough for real-time use; the paper itself notes that measured MA, FAS, and PASS datasets are scarce and sim-to-real gaps remain.","fun_headline_variants_meta":{"raw":{"variants":["AI unifies control of movable, fluid, and pinching antennas","One AI playbook for antennas that move, flow, and pinch","Survey: AI learns to reconfigure antennas with speed and adaptability","AI techniques compared across three reconfigurable antenna types","Spatial antenna flexibility meets AI for 6G wireless"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000756,"raw_usage":{"total_tokens":3205,"prompt_tokens":756,"completion_tokens":2449,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":2363}},"tokens_in":500,"tokens_out":2449,"duration_ms":19675,"temperature":1.0,"reasoning_tokens":2363,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T00:53:11.179332+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled hardware testbed would settle it: train any of the surveyed MA, FAS, or PASS policies on the standard simulation channels, then run them on measured channels in the same deployment. If the learned policy's throughput, secrecy rate, or sensing accuracy is no better than a fixed-position baseline once acquisition overhead and actuator/switching latency are included, the common-framework claim fails. The paper itself pinpoints this gap: most policies are trained and evaluated in simulation, so the test is whether the gains survive transfer.","supporting_citations":[],"review_version":1}