{"id":"69854f8d-94cc-4d4b-aa46-b6ee7095b7a9","arxiv_id":"2608.08033","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"WiFo-INR encodes partial CSI into compact modulation tokens that adapt a SIREN decoder, improving channel reconstruction and CSI feedback while cutting inference latency.","lead":"WiFo-INR is a wireless foundation model that represents channel state information as a neural function of time, frequency, and antenna coordinates instead of as a discrete grid. It shows gains in channel reconstruction and CSI feedback with lower latency, and it transfers to unseen CSI sizes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-instance modulation is the paper's central premise but is never isolated; a constant-token/shuffled-token control is needed to show that encoder-generated modulation tokens, rather than the shared SIREN decoder, drive the zero-shot gains.","rationale":"I agree with the reader's weakest_assumption: the cross-instance modulation path is the load-bearing element. My concrete test is not among the reader's four conditions, so I would keep CONDITIONAL and add it. I also note manuscript-level inconsistencies that should be fixed: Section V-D claims 'all ablations degrade' while Table XIII shows full training is 0.13 dB better on GC1-GC4; and the '3.27 dB' improvement in Section V-C1 is computed against the full-shot LLM baselines, not the strongest zero-shot baseline (WiFo-2), so the wording should be clarified. Error bars and INR baselines [22,23] remain useful but secondary; even with perfect numbers, the no-modulation control is needed to attribute the reported performance to the proposed mechanism.","tokens_in":20845,"tokens_out":10407,"duration_ms":110829,"concrete_test":"Evaluate WiFo-INR-Small (or Base) on the same pretraining and zero-shot splits with two controls: (1) replace the encoder output X_m in Eq. (13) with a single learned constant modulation vector shared by all instances; (2) at evaluation, randomly permute X_m across instances so each CSI is decoded with another instance's modulation. Compare average NMSE on PC/PF and GC/GF (Tables III–V and VII). If the constant/shuffled control degrades NMSE by less than about 1 dB, the encoder's instance-specific modulation is not the source of the claimed gains; if it degrades by several dB, the premise is confirmed. Run at least 3 seeds and report mean plus or minus standard deviation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central novelty is that a Transformer encoder generates instance-specific modulation tokens that adapt a shared SIREN decoder without per-instance optimization (Eq. (10), Eq. (13)). The strongest claims—zero-shot generalization to unseen CSI sizes and low-latency inference—all depend on this cross-instance modulation actually performing per-instance work. Yet no ablation or control isolates it. Table XII removes Fourier bases, raw coordinates, or replaces SIREN with MLP, but never disables or shuffles the modulation tokens. Because inputs are per-sample power-normalized and coordinates are normalized to [0,1] with a fixed Gaussian Fourier basis, a size-independent shared decoder could plausibly produce good reconstructions on its own; the reported GC/GF improvements over WiFo-2 would then be due to the SIREN decoder's inductive bias, not to the encoder's instance-adaptive mechanism. This is exactly the cross-instance modulation assumption the reader identified, and it is currently supported only by end-to-end metrics, not by any direct test.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WiFo-INR, a wireless foundation model that represents CSI as an implicit neural function. A Transformer encoder maps partial or coarse CSI to fixed-dimensional modulation tokens that modulate a SIREN-based decoder, avoiding per-instance optimization. A compression autoencoder quantizes the tokens for CSI feedback. The model is pretrained in two stages: mixed masking/denoising for 3D reconstruction, and compression-enhanced pretraining for 2D feedback. Experiments on 32 pretraining and 8 generalization datasets across channel estimation, time/frequency prediction, CSI feedback, and three downstream tasks claim improved NMSE, lower inference latency, and zero-shot generalization to unseen CSI sizes compared with WiFo-2, WiFo-CF, and task-specific baselines. Ablations and scaling analyses are provided.","tokens_in":21066,"tokens_out":4579,"duration_ms":46268,"significance":"If the empirical claims hold, WiFo-INR represents a meaningful step toward a generalizable, size-independent CSI representation with low inference cost, addressing a recognized limitation of token-based wireless foundation models. The paper ships a large synthetic benchmark suite and evaluates across several tasks, which is a strength. The central novelty—instance-adaptive modulation of a shared SIREN decoder without per-instance optimization—is timely and relevant. However, the core mechanism is not isolated by any ablation, and the absence of statistical error bars weakens several headline comparisons. The work is potentially significant for AI-native physical-layer design, but the evidence currently supports a qualified, not definitive, conclusion.","major_comments":[{"comment":"The central architectural premise is that the Transformer encoder's modulation tokens adapt the shared SIREN decoder per instance (Eq. (13)), and the paper attributes the zero-shot and low-latency gains to this cross-instance modulation. However, no ablation isolates this mechanism: Table XII removes Fourier bases, removes raw coordinates, and replaces SIREN with MLP, but never removes, shuffles, or replaces the modulation tokens. Without a control using (a) a constant, input-independent token and (b) shuffled per-instance tokens, the reported gains could plausibly originate from the SIREN decoder's sinusoidal inductive bias and the normalized-coordinate Fourier features alone. Please add these controls for at least the frequency-domain channel prediction and CSI feedback tasks, reporting NMSE on both pretraining and generalization splits. This is load-bearing because the abstract and conclusion explicitly attribute the advantages to encoder-generated instance-specific parameters.","section":"§IV-A, §V-D, Fig. 2"},{"comment":"No error bars, confidence intervals, or multiple-seed results are reported anywhere. Several headline comparisons rest on differences that are small relative to typical training noise: in Table XI, WiFo-INR's average F1 (0.761) exceeds WiFo-2 (0.753) by 0.008; in Table XIII, the default LoRA scheme on GC1–GC4 (−15.13 dB) differs from full-parameter training (−15.26 dB) by 0.13 dB. Without repeated runs or a statistical test, the claimed superiority over the strongest baselines cannot be distinguished from random seed variation. Please report mean and standard deviation over at least three seeds for the central results (channel reconstruction averages, CSI feedback averages, downstream-task metrics), or justify why single-run comparisons are conclusive in this setting.","section":"§V, Tables III–XIII"},{"comment":"The paper advertises 'zero-shot generalization to unseen CSI sizes' as a core advantage (Sec. III-B, fourth bullet; abstract). For channel reconstruction, the generalization datasets GC1–GC4 and GF1–GF4 are described in Table I and Sec. V-A as having unseen frequencies and scenarios, but the text does not identify which of their CSI sizes (e.g., T=16, K=64, N=8) fall outside the pretraining size distribution, nor does it evaluate a reconstruction task on such an unseen size in isolation. The only explicit size-transfer test is scenario classification (Table XI, S7). Please state explicitly which sizes are unseen during pretraining and provide a direct zero-shot evaluation for at least one reconstruction task on a size not present in the pretraining set, so the size-independence claim is empirically grounded.","section":"§V-C1, Table I, §III-B"}],"minor_comments":[{"comment":"The baselines WiFo-2 and WiFo-CF are the authors' own prior publications (Refs. [7], [13], [14]). While this is not inappropriate, the manuscript should disclose this more prominently in the experimental setup rather than only through the reference list, to help readers calibrate the comparisons.","section":"§V-B2"},{"comment":"The Fourier features in Eq. (11) use a Gaussian matrix B with 'domain-specific variances' σ_t^2, σ_s^2, σ_f^2, but Sec. V-B1 states all three variances are set to 1. Please clarify whether domain-specific values are used anywhere or report the actual settings.","section":"Eq. (11), §V-B1"},{"comment":"For datasets with compression ratio 1/16, WiFo-CF is listed as '/', because WiFo-CF is pretrained at a fixed ratio of 1/32. Please state this explicitly in the table caption or text so the missing entries are not mistaken for unreported results.","section":"Table VII"},{"comment":"The sentence 'reducing the average NMSE by 3.28 dB and 5.63 dB compared with TransNet and CsiNet' refers to full-shot performance on the pretraining datasets, but the preceding paragraph also discusses zero-shot performance. Please separate the full-shot and zero-shot claims to avoid ambiguity.","section":"§V-C2"},{"comment":"In Table VI, the 'Transformer' baseline has 1.08M parameters; since it is trained per dataset, please clarify whether the parameter count is for a single dataset or an average, and how the inference time is measured on comparable hardware.","section":"§V-C1, Table VI"},{"comment":"The notation for the temporally padded CSI in the compression pretraining stage, 'fHCF' in Sec. IV-B2, is inconsistent with the tilde notation used elsewhere; please use e.g. '\\tilde{H}_{CF}' throughout.","section":"§IV-B2, Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on the authors' own prior works (WiFo, WiFo-2, WiFo-CF) as baselines and sources of pretraining tasks. This is not disqualifying, but it should be declared in the paper. The synthetic-data evaluation is broad but would benefit from at least one real-world or standardized dataset in addition to the DICHASUS localization experiment. The missing modulation-token ablation and lack of error bars are the main technical obstacles to acceptance; both are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things up front. First, this is the first cross-instance generalizable INR for wireless CSI that I have seen: a Transformer encoder produces modulation tokens that adapt a shared SIREN decoder, avoiding the per-instance optimization of earlier INR work. Second, the empirical sweep is unusually broad: 32 pretraining datasets, 8 held-out generalization datasets, three reconstruction tasks, CSI feedback, three downstream tasks, and a scaling analysis. The zero-shot results, often beating full-shot baselines, are impressive if true.\n\nWhat the paper does well is real. The two-stage pretraining (masking/denoising for 3D reconstruction, then compression-enhanced for 2D feedback) is sensible and the ablations in Table XII show the SIREN decoder and Fourier features matter. The scaling laws are a nice touch. The inference-latency reductions versus Transformer decoders are concrete and plausible given the lightweight SIREN decoder.\n\nThe soft spots, in proportion. The largest is that the paper's central premise—that the encoder's instance-specific modulation tokens are what drive the gains—is never directly tested. Table XII removes Fourier bases, raw coordinates, and swaps SIREN for MLP, but it never disables or shuffles the modulation tokens. Because inputs are per-sample normalized and coordinates are fixed, one could imagine a shared SIREN decoder doing most of the work on its own. A constant-token or shuffled-token control would settle this. This is not fatal, but it is a real gap in the causal story.\n\nSecond, there are no error bars or multiple seeds anywhere. The reported advantages are often 1–3 dB, and against the strongest baselines some of those gaps could be within run-to-run noise. Third, the two most direct INR priors ([22] and [23]) are cited but not included as baselines, even for the tasks where they are directly applicable. Fourth, no code or data is released, which is normal for this subfield but still limits trust. Finally, Table XIII contains a minor overclaim: on the GC split, \"Full training\" actually beats the default scheme slightly, though with far more trainable parameters; the text calls them comparable, which is fair, so this is cosmetic.\n\nBottom line: this paper deserves a serious referee. The architecture is novel, the experiments are extensive, and the limitations are fixable. I would send it to review with a conditional accept-comment list: add the modulation control, report error bars or seeds, include the INR baselines, and release code or at least detailed implementation. The paper is for wireless foundation-model researchers and anyone working on INR generalization; it is worth their time.","headline":"A genuinely new INR-based wireless foundation model with broad empirical support, but the central modulation mechanism needs an isolating control and the quantitative claims need error bars.","tokens_in":21598,"tokens_out":1876,"would_cite":true,"duration_ms":20912,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"WiFo-INR models channel state information as a coordinate-conditioned neural function, claiming this implicit representation outperforms discrete-tensor wireless foundation models on channel reconstruction and CSI feedback while cutting…","keywords":["implicit neural representation","wireless foundation model","channel state information","channel reconstruction","CSI feedback","SIREN","self-supervised pretraining","zero-shot generalization"],"falsifier":"Feed WiFo-INR a set of CSI tensors from a propagation environment with markedly different statistics, such as a different carrier frequency or a non-stationary scenario; record the encoder-generated modulation tokens and the reconstructed NMSE, and compare against the same decoder with per-instance gradient-optimized modulation tokens. If the encoder's tokens cannot match or approximate the per-instance optimum, the cross-instance modulation premise underlying zero-shot generalization fails.","tokens_in":20642,"feed_emoji":"📡","tokens_out":5971,"duration_ms":60409,"temperature":0.7,"pith_summary":"The paper sets out to replace the discrete-tensor treatment of channel state information in wireless foundation models with an implicit neural representation: a SIREN decoder that maps time-frequency-space coordinates to complex channel values, adapted per instance by modulation tokens produced by a Transformer encoder. It argues that this coordinate-conditioned design captures the multipath-induced oscillatory structure of CSI more efficiently than generic token decoders, produces fixed-size instance representations independent of CSI dimensions, and removes per-instance optimization. If the claims hold, a single pretrained wireless model can reconstruct channels, feed compressed CSI back, and transfer to localization, beam prediction, and scenario classification with lower inference latency and far less fine-tuning overhead than current foundation models.","feed_headline":"3.28 dB better CSI feedback with a neural-function channel model","feed_subtitle":"WiFo-INR's SIREN decoder rebuilds channels from partial CSI, generalizes to unseen sizes, and cuts inference latency.","key_machinery":"The load-bearing mechanism is the pair consisting of encoder-generated modulation tokens and a SIREN decoder. SIREN is a multilayer perceptron with sinusoidal activations; because a CSI tensor is a superposition of multipath sinusoids, this provides a frequency-aware inductive bias. The Transformer encoder maps partial or coarse CSI to $r$ fixed-dimensional tokens $\\mathbf{X}_m$, and those tokens modulate only the first SIREN layer through $\\mathbf{h}_0 = \\sin(\\omega_0(\\mathbf{U}\\mathbf{X}_m\\mathbf{h}_f + \\mathbf{b}_0))$, so the same shared decoder can represent many instances without per-instance optimization. Normalized space-time-frequency positional encoding makes the tokens independent of CSI size, and a compression autoencoder reduces the token count and dimension for quantized feedback.","core_discovery":"The central claim is that a generalizable implicit neural representation architecture, called WiFo-INR, can serve as a wireless foundation model. The model represents each CSI instance as a neural function $f_{\\psi,\\phi}(s,k,n)=\\hat{H}_{s,k,n}$, where $\\psi$ are shared decoder weights and $\\phi$ are instance-specific modulation tokens generated by an encoder from partial or coarse CSI, eliminating online gradient-based per-instance optimization. The decoder uses sinusoidal activations and Fourier features, giving it a frequency-aware inductive bias matched to multipath channel structure. A two-stage self-supervised pretraining scheme first learns 3D channel reconstruction through mixed masking and denoising, then adapts the model for quantized 2D CSI feedback through compression-enhanced pretraining. In the reported experiments the model reduces average NMSE on generalization datasets by 3.27 dB across three reconstruction tasks, lowers CSI feedback NMSE by 3.28 dB over TransNet, achieves zero-shot feedback performance 2.36 dB better than TransNet's full-shot performance, and cuts inference time by 44.56 percent relative to WiFo-2.","pith_inferences":["Beyond the paper's results, the same encoder-to-modulation design should transfer to other radio-frequency map reconstruction problems, such as radio environment mapping or sensing-aided channel estimation, where signals are also smooth functions of coordinates.","The size-independence property implies a practical deployment path: one pretrained model can serve devices with different antenna arrays and bandwidths without retraining heads, a property the paper demonstrates on scenario classification but does not pursue for other tasks.","A testable extension would be to add a few gradient steps of per-instance optimization at inference time; if the encoder already produces near-optimal modulation tokens, the extra steps should yield only marginal NMSE gains, which would directly quantify how much of the performance is due to the cross-instance mechanism."],"forward_implications":["A single pretrained WiFo-INR can reconstruct channels from pilots, predict channels in time and frequency, and compress CSI for feedback, replacing several task-specific models.","Because modulation tokens are fixed-dimensional regardless of CSI size, the same model and output head can be applied to unseen antenna, subcarrier, and temporal configurations without redesign.","Zero-shot performance on unseen CSI sizes and scenarios beats the full-shot performance of task-specific baselines in the reported experiments, so deployment to new environments could skip per-scenario retraining.","The lightweight SIREN decoder lowers inference latency relative to Transformer decoders, making the model more feasible on resource-constrained user equipment.","The two-stage pretraining lets one set of weights handle both 3D channel acquisition and 2D feedback, with the second stage using LoRA to keep trainable parameters low."],"supporting_citations":[{"why":"Supplies the SIREN architecture with sinusoidal activations that the decoder uses as the implicit neural representation.","marker":"[20]"},{"why":"Supplies the generalizable INR mechanism of generating instance-specific modulation parameters from an encoder, removing per-instance optimization.","marker":"[25]"},{"why":"Provides early evidence for representing MIMO CSI as a neural function and motivates the feedback-oriented INR design.","marker":"[22]"},{"why":"Supplies the WiFo-CF baseline and the mu-law quantized feedback protocol used by the compression autoencoder.","marker":"[13]"},{"why":"Supplies the WiFo-2 baseline, the mixed-masking and denoising pretraining tasks, and the downstream-task comparisons.","marker":"[14]"},{"why":"Supplies Fourier features that let the coordinate decoder fit high-frequency channel variations.","marker":"[26]"},{"why":"Supplies the standardized channel simulation dataset used to build the pretraining and generalization corpora.","marker":"[27]"}],"fun_headline_variants":["Neural function CSI: 3.28 dB better feedback, 44.56% faster","WiFo-INR uses implicit neural reps to cut inference latency by 44.56%","Zero-shot CSI to unseen sizes with INR-based wireless foundation model","SIREN decoder lifts CSI feedback 3.28 dB and slashes latency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that a Transformer encoder can produce, from partial or coarse CSI, modulation parameters that adapt one shared SIREN decoder to reconstruct any CSI instance, including instances from unseen environments, frequency bands, and CSI sizes, is what carries the zero-shot and low-latency claims.","fun_headline_variants_meta":{"raw":{"variants":["Neural function CSI: 3.28 dB better feedback, 44.56% faster","WiFo-INR uses implicit neural reps to cut inference latency by 44.56%","Zero-shot CSI to unseen sizes with INR-based wireless foundation model","SIREN decoder lifts CSI feedback 3.28 dB and slashes latency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1290,"prompt_tokens":982,"completion_tokens":308,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":220}},"tokens_in":598,"tokens_out":308,"duration_ms":4426,"temperature":1.0,"reasoning_tokens":220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:31:17.838205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed WiFo-INR a set of CSI tensors from a propagation environment with markedly different statistics, such as a different carrier frequency or a non-stationary scenario; record the encoder-generated modulation tokens and the reconstructed NMSE, and compare against the same decoder with per-instance gradient-optimized modulation tokens. If the encoder's tokens cannot match or approximate the per-instance optimum, the cross-instance modulation premise underlying zero-shot generalization fails.","supporting_citations":[{"cited_title":"Implicit Neural Representations with Periodic Activation Functions,","cited_arxiv_id":null,"evidence_quote":"Supplies the SIREN architecture with sinusoidal activations that the decoder uses as the implicit neural representation."},{"cited_title":"Gener- alizable Implicit Neural Representations via Instance Pattern Composers,","cited_arxiv_id":null,"evidence_quote":"Supplies the generalizable INR mechanism of generating instance-specific modulation parameters from an encoder, removing per-instance optimization."},{"cited_title":"WiFo-CF: Wireless Foundation Model for CSI Feedback,","cited_arxiv_id":null,"evidence_quote":"Supplies the WiFo-CF baseline and the mu-law quantized feedback protocol used by the compression autoencoder."},{"cited_title":"Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains,","cited_arxiv_id":null,"evidence_quote":"Supplies Fourier features that let the coordinate decoder fit high-frequency channel variations."},{"cited_title":"QuaDRiGa: A 3-D Multi-Cell Channel Model With Time Evo- lution for Enabling Virtual Field Trials,","cited_arxiv_id":null,"evidence_quote":"Supplies the standardized channel simulation dataset used to build the pretraining and generalization corpora."}],"review_version":1}