{"id":"097bcae9-9842-42ea-95a5-ac45f10e90d5","arxiv_id":"2509.06968","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A systematic review of deep learning techniques for integrated sensing and communication (ISAC), spanning waveform design, beamforming, channel estimation, and receivers.","lead":"This paper surveys how deep learning is being used to build wireless systems that communicate and sense their surroundings at the same time. It catalogs dozens of recent methods, compares their reported speed and accuracy, and lays out open problems for 6G networks.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 7's complexity claims rest on non-comparable execution times from different hardware and problem sizes; the survey admits direct comparison is difficult yet generalizes that DL complexity 'slightly increases' with antennas while iterative methods 'exponentially increase.'","rationale":"The reader's weakest_assumption identified that the survey's complexity comparisons rely on non-comparable numbers from different simulators, hardware, and problem sizes. I agree; this is the most load-bearing concern because the abstract's claim of 'reduced computational complexity' is central to the paper's thesis, and Section 7 is the only direct support. The paper explicitly acknowledges the difficulty of comparison, yet still makes specific scaling claims ('slightly increases' vs 'exponentially increase') without any formal analysis. This is not a matter of external consensus; it is an internal gap between the evidence cited and the generalization made. A concrete test—deriving FLOPs scaling or running a controlled benchmark—would settle whether the claim lands. Since the paper is a survey with no new experiments, and the central claim is not falsifiable from the text alone, the appropriate verdict remains UNVERDICTED. My concern does not change the reader's verdict; it sharpens the rationale for why the complexity claim is unverified.","tokens_in":40523,"tokens_out":5235,"duration_ms":58605,"concrete_test":"Derive the computational complexity (FLOPs per inference) of the SLP deep learning method in [182] as a function of the number of antennas N, and compare it with the per-iteration complexity of the iterative algorithms in [181] (penalty dual decomposition and Riemannian BFGS). If the DL network's FLOPs scale superlinearly (e.g., O(N^2) or worse due to hidden layer dimensions) while the iterative methods scale polynomially (e.g., O(N^3) per iteration), then the survey's claim that DL complexity 'slightly increases' with antennas while iterative methods 'exponentially increase' is unsupported. A complementary empirical check: run the released code (if available) of both methods on the same hardware with N = 16, 32, 64, 128 and measure the scaling exponent of wall-clock time. If the DL exponent is not significantly lower than the iterative exponent, the central complexity advantage is not v","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim that DL-based techniques offer 'reduced computational complexity' and are suitable for 'limited computational resources' rests primarily on Section 7. Yet Section 7 contains an internal tension: it states 'it is difficult to provide a direct computational complexity comparison between the DL-based techniques and optimization-based or iterative methods' (paragraph after Table 10), then proceeds to make broad, unquantified generalizations. Specifically, Section 7.2 claims DL execution times are 'microseconds to milliseconds' and that DL execution time 'slightly increases with the number of antennas' while iterative methods 'may exponentially increase.' The only concrete evidence cited is one example: SLP precoding from [181] (9.91 s and 0.233 s) versus [182] (0.01–0.02 s). These are wall-clock times from different studies, almost certainly on different hardware, with different problem sizes and implementations. No scaling analysis is provided to justify the antenna-scaling claim. This is load-bearing because the abstract's complexity advantage is the primary motivation for deploying DL in real-time ISAC. If, for example, DL inference complexity scales linearly or worse with antenna count (e.g., due to dense layers whose input size grows with N), while iterative methods scale polynomially (e.g., O(N^3) per iteration), then the survey's claim that DL complexity 'slightly increases' while iterative methods 'exponentially increase' would be misleading. The paper itself acknowledges the difficulty of direct comparison, underscoring that the stated complexity advantage is not established by the cited evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a survey of deep learning (DL) techniques for integrated sensing and communication (ISAC) systems. It reviews DL learning strategies and architectures, ISAC fundamentals and performance metrics, and then categorizes recent work on DL-based waveform design, beamforming/precoding, predictive beamforming, channel estimation, and receiver processing (demodulation, target estimation, gesture recognition). Section 7 addresses computational complexity, arguing that DL inference is markedly faster than iterative/optimization methods and scales more favorably with antenna count. Section 8 outlines challenges and future directions, including model-based DL, scalability, lightweight models, and networking. The abstract's central claim is that DL-based techniques provide near-optimal solutions with reduced computational complexity, making them suitable for real-time, resource-limited ISAC systems.","tokens_in":40889,"tokens_out":5629,"duration_ms":58709,"significance":"If the survey is reliable, it provides a useful structured reference for the ISAC and machine-learning communities. Its strengths include a clear taxonomy of DL architectures and their ISAC use cases, comparative tables across many recent works (including 2023–2025 papers), and a dedicated treatment of training and inference complexity. The paper also identifies open challenges that are generally well chosen. However, the load-bearing complexity claim in Section 7 rests on non-comparable wall-clock times and is internally inconsistent with statements elsewhere in the manuscript. These issues must be addressed before the survey's central thesis can be considered supported.","major_comments":[{"comment":"The manuscript acknowledges 'it is difficult to provide a direct computational complexity comparison between the DL-based techniques and optimization-based or iterative methods' (paragraph after Table 10), but then makes broad quantitative assertions: iterative methods 'can take up to 10 s or more' and DL techniques 'can be executed in a few hundred milliseconds,' and that iterative methods 'may exponentially increase' with antennas while DL 'slightly increases.' The only concrete evidence is one SLP example ([181] vs. [182]) with wall-clock times from different studies, almost certainly on different hardware and problem sizes. Table 11 lists generic ranges (0.1–30 s vs. microseconds–milliseconds) with no sources or experimental conditions. Because the abstract's central claim of 'reduced computational complexity' is supported primarily by these assertions, they need to be either rigorou","section":"Section 7.2, Table 11"},{"comment":"There is a direct internal contradiction. Table 9, row [43], lists 'Exponentially increasing complexity with the number of antennas' as a note for a DNN-transformer receiver. Section 7.2, however, states that 'the execution time of DL-based techniques slightly increases with the number of antennas.' These two claims cannot both be true for the surveyed literature. Moreover, Section 8.1 and Section 8.3 state that DL models are trained for a fixed number of inputs/outputs and that changing the number of antennas may require retraining, which is also inconsistent with a general claim of benign scaling with antenna count. The authors should either restrict the scaling claim to specific architectures (e.g., fully connected networks with fixed input dimension) or explicitly acknowledge counterexamples such as [43].","section":"Table 9, Section 7.2"},{"comment":"The methodological basis for the complexity comparison is thin. Table 10 gives per-layer real-multiplication counts for generic NN architectures, but the text then jumps to categorical execution-time comparisons without explaining how per-layer counts translate to end-to-end ISAC task complexity, training cost, memory usage, or parallelization benefits. The table also omits definitions for all symbols (NN, NI, NF, NK, NO, NS, NH—the duplicate 'NI' is confusing; presumably one is NH). Since the section is the paper's principal evidence for its central claim, the authors should either connect the per-layer complexity model to the cited ISAC applications or clearly frame the Section 7.2 statements as qualitative observations from individual papers rather than general laws.","section":"Section 7, Table 10"}],"minor_comments":[{"comment":"The first term inside the expectation should be R(i0), not i0; as written it is dimensionally inconsistent with the reward notation. Please check the transcription.","section":"Equation (4)"},{"comment":"The notation 'γPT_{t=1} ψ_t(W)' appears to be a typesetting error for a summation over t (likely ∑_{t=1}^T ψ_t(W)). Please correct and define T.","section":"Equation (6)"},{"comment":"In the introduction, 'Table 3 1 presents the summary' appears to be a typo; it should likely read 'Table 1' or 'Table 3.' Check the cross-reference.","section":"Section 1"},{"comment":"The sentence 'when the number of training samples is limited, it is performance is degraded' contains a grammar error; also the DNN channel-estimator claim about outperforming LMMSE only with high training samples is stated a bit ambiguously.","section":"Section 5.2"},{"comment":"The Notes column says 'Extremely training compared to CNN-based estimators,' which is ungrammatical. It presumably means 'Extremely fast training' or 'Extremely low training time.'","section":"Table 7, row [39]"},{"comment":"Typo: 'such as ash mobile devices' should be 'such as mobile devices' (or 'smartphones'). Also, Section 8.9 contains a duplicate sentence about semantic communication in ISAC systems; please remove the repetition.","section":"Section 8.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a survey by authors who are active in the ISAC/DL field, and the self-citations are relevant to the sections in which they appear. The main obstacle is not novelty or scope but the rigor and internal consistency of Section 7, which directly supports the abstract's headline claim. If the authors can revise the complexity discussion to be properly qualified and reconcile the contradictions noted above, the survey could be suitable for publication. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent and current survey of deep learning for ISAC, and someone new to the area would come away with a good map of the literature. The one part I would not take at face value is Section 7's computational complexity discussion.\n\nWhat it does well: the paper organizes a lot of material. Table 3 gives a taxonomy by application area (waveform design, beamforming, channel estimation, receiver processing), and the per-section tables summarize each cited method's learning strategy, architecture, inputs, outputs, pros, and caveats. It covers recent 2024-2025 papers that earlier surveys miss, and the background on learning strategies and DL architectures is accurate and readable. For a survey, that is real value.\n\nWhere it is soft: Section 7 argues that DL-based methods have much lower inference complexity than iterative or optimization-based methods. The paper itself says \"it is difficult to provide a direct computational complexity comparison,\" then goes ahead and generalizes that DL runs in microseconds to milliseconds, that its execution time \"slightly increases\" with antennas, and that iterative methods \"may exponentially increase.\" The only concrete evidence is one SLP example: 9.91 s and 0.233 s for two iterative algorithms versus 0.01-0.02 s for a DL approach. Those are wall-clock numbers from different papers, almost certainly on different hardware, simulators, and problem sizes. No scaling analysis backs the antenna-dependence claim. The qualitative direction is probably right—DL inference does tend to be faster than iterative optimization at runtime—but the survey overstates it as a settled quantitative fact. This matters because the abstract's claim about \"reduced computational complexity\" leans on that section. A revised version should either add a controlled comparison, cite one if it exists, or hedge the generalization more carefully.\n\nMinor issues: the survey does not disclose how papers were selected for inclusion, which is common but worth noting; and there are a few typos and sloppy equations (e.g., the discount factor in Eq. (4)).\n\nOverall: the paper is a legitimate survey, not a new result, and the central tension in Section 7 is a moderate flaw rather than a fatal one. The taxonomy and coverage make it worth a serious referee; I would send it out but ask for Section 7 to be rewritten with proper caveats.","headline":"A competent, current survey of DL-for-ISAC that delivers a useful taxonomy; the computational complexity claims in Section 7 go beyond what the cited evidence supports and need a rewrite.","tokens_in":41326,"tokens_out":2311,"would_cite":true,"duration_ms":25472,"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":"This survey argues that offline-trained neural networks, used as learned surrogates for non-convex optimization, are the practical path to real-time integrated sensing and communication, replacing iterative solvers whose execution times rea","keywords":["deep learning","integrated sensing and communication","6G networks","waveform design","channel estimation","beamforming","computational complexity","real-time signal processing"],"falsifier":"Run a controlled benchmark on identical hardware where an iterative ISAC solver (for example, the 0.233-second augmented-Lagrangian method of [181]) and a DL surrogate are tested on the same antenna count, channel model, and SNR range; if the DL model's inference time exceeds the iterative solver's, or its sum rate and detection probability fall materially below the advertised near-optimal values, the survey's general complexity-reduction claim fails.","tokens_in":40494,"feed_emoji":"📡","tokens_out":5879,"duration_ms":69607,"temperature":0.7,"pith_summary":"Integrated sensing and communication (ISAC) lets one waveform and one hardware platform both deliver data and sense the environment, a feature 6G networks are expected to rely on; its obstacle is the heavy iterative and optimization-based signal processing needed to design waveforms, beamformers, channel estimates, and receivers. This survey tries to establish that deep learning can remove that obstacle: trained networks can produce near-optimal solutions across all these ISAC modules while running in microseconds to milliseconds, because the expensive training happens offline and inference is only matrix multiplication. The payoff, if the survey is right, is that ISAC becomes deployable on low-power, latency-constrained devices, and the same hardware can handle sensing and communication without the computational burden that currently forces a trade-off between the two. The survey supports this by categorizing dozens of recent studies and comparing their reported runtimes and performance against conventional baselines.","feed_headline":"Survey: deep learning makes 6G sensing-plus-communication fast enough","feed_subtitle":"Review of recent DL-based ISAC work shows trained networks can replace slow iterative solvers in 6G.","key_machinery":"The load-bearing mechanism is the trained neural network as a learned surrogate for a hard optimization problem. In the surveyed systems, a network maps raw inputs (CSI, pilot symbols, received waveforms, echo vectors) directly to the desired solution (a waveform matrix, precoder, channel estimate, demodulated bits, or target range/velocity/angle). The property carrying the argument is the separation of training from inference: training cost is paid offline on powerful hardware, while inference only performs matrix and vector multiplications that scale gently with the number of antennas and can run in parallel on low-power accelerators. The architectures reviewed include FCDNN/MLP, CNN, LSTM","core_discovery":"The paper's central claim is that DL-based techniques are a viable, efficient alternative to conventional iterative or optimization-based methods for ISAC system design. It asserts that trained neural networks supply near-optimal solutions for waveform optimization, symbol-level precoding, channel estimation, data demodulation, and target parameter estimation, with inference times in the microsecond-to-millisecond range versus 0.1–30 seconds for the surveyed optimization-based methods. The evidence is organized by system module: transmitter-side waveform and precoder design, channel estimation, and receiver-side demodulation and sensing processing. Across these categories the survey highligh","pith_inferences":["The paper compares runtimes across different simulators and hardware rather than through a controlled benchmark; if such a benchmark reproduced the reported gap, the central complexity claim would be confirmed on firmer ground.","The learned-surrogate pattern likely extends beyond the surveyed modules to joint localization-and-communication and semantic ISAC, since those tasks share the same structure of optimizing under a dual objective—the paper only hints at these as future directions.","A testable consequence is that a quantized CNN or LSTM trained for a fixed antenna count should need retraining when the antenna count changes; the paper names scalability as a gap, implying that fixed-size architectures will not gracefully generalize.","The field would benefit from standard ISAC datasets and hardware baselines; absent those, the near-optimality claim remains a collection of per-paper demonstrations rather than a verified universal property."],"forward_implications":["ISAC transmitters could update precoders and waveforms within a few hundred milliseconds, tracking changing channel conditions that multi-second iterative solvers cannot keep up with.","Channel estimators trained offline can beat LS and LMMSE in ISAC-specific interference scenarios, and extreme learning machines can train in a fraction of the time of gradient-based networks while keeping comparable accuracy.","Receiver-side DL can jointly demodulate communication data and estimate sensing parameters, in some cases removing the need for separate interference-cancellation stages.","Weight quantization and pruning can shrink a DL-based ISAC precoder's memory and compute to about one-sixth while keeping most of its performance, making edge-device deployment plausible.","Federated and transfer learning can reduce training time and data requirements, enabling multi-cell ISAC networks to cooperate without sharing raw sensing data."],"supporting_citations":[{"why":"Supplies evidence that model-based online reinforcement learning can optimize ISAC waveforms with fewer samples and fast convergence, replacing non-convex optimization.","marker":"[20]"},{"why":"Shows the model-free data-driven counterpart, establishing the trade-off between sample efficiency and asymptotic performance that the survey uses to frame DL-based waveform optimization.","marker":"[21]"},{"why":"Demonstrates a low-complexity unsupervised-learning ISAC precoder where quantization and pruning retain 96.1% of sum rate with 16.88% of memory and compute; a core evidence point for the complexity-reduction claim.","marker":"[27]"},{"why":"Provides the ELM-based channel estimator that trains about five times faster than an FCDNN with comparable estimation accuracy, supporting the training-complexity comparison.","marker":"[39]"},{"why":"Supplies a THz ISAC receiver where two neural networks demodulate data and estimate target range and velocity with lower computational complexity than MUSIC.","marker":"[40]"},{"why":"Offers a transformer-based ISAC receiver that jointly estimates data symbols and target angle/time-delay and outperforms other DL and LS demodulators, supporting receiver-side gains.","marker":"[43]"},{"why":"Demonstrates vertical federated edge learning for distributed ISAC sensing, supporting the survey's claims on privacy, communication overhead, and network-level cooperation.","marker":"[48]"},{"why":"Supplies a deep-unfolding beamforming method that reduces complexity relative to optimization, backing the model-based DL direction.","marker":"[180]"},{"why":"Gives the iterative symbol-level precoding baselines with execution times of 9.91 s and 0.233 s, the direct comparison points for DL inference speed.","marker":"[181]"},{"why":"Provides a DL-based symbol-level ISAC precoder with execution times of 0.01–0.02 s, the key evidence for the survey's real-time inference claim.","marker":"[182]"}],"fun_headline_variants":["Deep learning accelerates 6G sensing and communication","Neural nets replace slow iterative ISAC solvers","DL cuts ISAC computing from seconds to milliseconds","How trained networks speed up 6G ISAC systems","Deep learning-enabled ISAC: fast and near-optimal"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The survey assumes that the runtime and performance numbers reported by different papers, produced on different simulators, hardware platforms, and problem sizes, can be compared directly to conclude that deep learning is generally fast and near-optimal for ISAC.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning accelerates 6G sensing and communication","Neural nets replace slow iterative ISAC solvers","DL cuts ISAC computing from seconds to milliseconds","How trained networks speed up 6G ISAC systems","Deep learning-enabled ISAC: fast and near-optimal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000812,"raw_usage":{"total_tokens":3413,"prompt_tokens":776,"completion_tokens":2637,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":2575}},"tokens_in":520,"tokens_out":2637,"duration_ms":21923,"temperature":1.0,"reasoning_tokens":2575,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:58:26.934273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled benchmark on identical hardware where an iterative ISAC solver (for example, the 0.233-second augmented-Lagrangian method of [181]) and a DL surrogate are tested on the same antenna count, channel model, and SNR range; if the DL model's inference time exceeds the iterative solver's, or its sum rate and detection probability fall materially below the advertised near-optimal values, the survey's general complexity-reduction claim fails.","supporting_citations":[{"cited_title":"Low-complexity joint radar- communication beamforming: From optimization to deep unfolding,","cited_arxiv_id":null,"evidence_quote":"Supplies a deep-unfolding beamforming method that reduces complexity relative to optimization, backing the model-based DL direction."},{"cited_title":"Dual- functional radar-communication waveform design: A symbol-level precoding approach,","cited_arxiv_id":null,"evidence_quote":"Gives the iterative symbol-level precoding baselines with execution times of 9.91 s and 0.233 s, the direct comparison points for DL inference speed."},{"cited_title":"SLP- based dual-functional waveform design for ISAC sys- tems: A deep learning approach,","cited_arxiv_id":null,"evidence_quote":"Provides a DL-based symbol-level ISAC precoder with execution times of 0.01–0.02 s, the key evidence for the survey's real-time inference claim."}],"review_version":1}