{"id":"3a69bbd6-9893-4bf3-b597-55370ac7f1f6","arxiv_id":"2509.01875","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A diffusion model reconstructs radio maps from sparse vertex-level RSS samples, yielding non-line-of-sight emitter localization without known transmit power.","lead":"RadioDiff-Loc uses a conditional diffusion model to reconstruct a full radio signal map from sparse signal strength readings taken at building edges and corners, then locates a hidden transmitter by finding the brightest point of that map. It claims localization with under 1% of positions sampled and no knowledge of the transmitter's power, which matters for emergency response and autonomous systems in obstructed environments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Power-invariant localization claim is untested and possibly inconsistent: Eq. (31) normalizes only sparse RSS inputs by max sampled value while the training target is absolute RadioMapSeer pathloss at fixed 23 dBm; no experiment varies transmitter power.","rationale":"Reader's weakest assumption (FI/MI w.r.t. u not d) is a genuine gap, and I partially agree with the reader: they also list normalization ambiguity as issue (3). But I would elevate the power-invariance issue over the sampling-theory issue as the single most load-bearing concern. The paper's own data show the learned pipeline localizes well even with random sampling at 0.96% (3.035 m, Table II), and the vertex advantage over random at matched budget is ~9% (Tables III/IV); so a failure of the sampling-theory link would weaken the 'physics-guided' narrative but not the core feasibility claim. In contrast, the claim 'without knowledge of transmit power' is in the abstract, introduction, and system design, and it rests entirely on Eq. (31). The evaluation never varies transmit power. Since the training target normalization is unspecified and the training data are fixed-power, the model may have memorized the absolute scale of RadioMapSeer; the power-invariance claim is therefore unsupported and potentially inconsistent. The proposed concrete test directly settles this. This does not change the reader's CONDITIONAL verdict, but it sharpens the condition: before accepting the claim, the authors must demonstrate power invariance under held-out transmit powers and specify the exact normalization applied to both inputs and targets.","tokens_in":16186,"tokens_out":8299,"duration_ms":100898,"concrete_test":"Take the RadioMapSeer test maps and simulate unknown transmitter power by adding a global offset Delta in {-20,-10,+10,+20} dB to the ground-truth pathloss maps (or multiplying linear-power maps by 10^{Delta/10}). Keep the sampling masks fixed, normalize the sparse RSS inputs by their sampled maximum as in Eq. (31), and run the published inference pipeline. Measure localization error and RMSE/SSIM of the reconstructed map. To disambiguate target-normalization ambiguity, run two variants: (a) the implementation exactly as described (absolute pathloss target); (b) with both the sparse inputs and the full training target normalized by their respective true map maxima. If localization error or RMSE changes systematically with Delta under both variants, the power-invariance claim fails; if LE is unchanged in at least one variant, the claim can be repaired by specifying that normalization proc","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes localization 'without knowledge of absolute transmission power' (Sec. IV-B). The only mechanism is Eq. (31): tilde{y}_i = y_i / max_j y_j, which the text says converts the RM into a relative pathloss map. But Sec. V-B states the network is trained with an MSE loss against the RadioMapSeer ground-truth pathloss maps, where every transmitter is at 23 dBm, and no normalization of the training target is specified. If the target is absolute pathloss, then at inference with a different transmit power the normalized inputs are identical to those at 23 dBm but the correct output map is shifted by a constant; the model cannot infer the missing scale. If instead the target is normalized by the true map maximum, then normalizing by max sampled RSS is a biased estimator of that maximum, since the source and its true peak lie in the inaccessible restricted region S_r and are never sampled. The training and inference conditionings would therefore be mismatched. The paper performs no experiment with varying transmit power, so the 'power-invariant' property is asserted but not demonstrated. This is more load-bearing than the sampling-theory gap: the same-budget comparisons in Table I show only mild gains for vertex/edge sampling over random (2.770 vs 3.035 m at 0.96%), whereas the unknown-power claim is a headline contribution and is entirely unvalidated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RadioDiff-Loc, a framework for non-line-of-sight emitter localization that reconstructs a dense radio map from sparse RSS measurements using a conditional diffusion model. The sampling locations are chosen near building edges/vertices, motivated by knife-edge diffraction theory; RSS inputs are normalized by the maximum sampled value to remove dependence on unknown transmit power; and the emitter is localized as the maximum-intensity point of the reconstructed map. The method is evaluated on the RadioMapSeer dataset and compared against classical RSS localization algorithms and alternative sampling strategies. The paper claims high localization accuracy with under 1% sampling and power-invariant inference.","tokens_in":16503,"tokens_out":7902,"duration_ms":85577,"significance":"If the claims were fully validated, the work would be a useful contribution: it combines physically motivated sparse sampling with generative radio-map reconstruction, addresses the unknown-power problem, and demonstrates a dual-driven use of reconstructed maps with classical post-processing. The use of a public benchmark dataset (RadioMapSeer) is a strength, as is the explicit attempt to connect sampling design to diffraction physics. However, the current validation does not support the headline claims: the power-invariance mechanism is not demonstrated, the main comparison against classical methods is confounded by training advantage and differing sampling budgets, and the theoretical justification for vertex sampling is not actually about the source position. These issues are load-bearing for the central claims, so the manuscript needs a substantive revision.","major_comments":[{"comment":"The claim that 'when the sampling budget is constrained, random sampling experiences an increase in localization error exceeding 50%, whereas vertex-based sampling shows a more graceful degradation with only a 40% increase' appears without a corresponding experiment. Table I contains only one sampling rate per strategy and no sweep of sampling budgets. This unsupported quantitative comparison is used to bolster the central sampling-efficiency claim and should be removed or substantiated with the relevant experiments.","section":"Sec. V-E, Table I"}],"minor_comments":[{"comment":"Typo: 'Scattering Congnition' should be 'Scattering Cognition'.","section":"Title, Abstract"},{"comment":"Typo: 'The forward diffusion process is described by the SDE ad follows' should be 'as follows'.","section":"Sec. II-B"},{"comment":"Equation numbering is inconsistent: Eq. (18) appears twice, Eq. (19) appears twice, and the localization estimator is labeled Eq. (21) despite no Eq. (20) in that section.","section":"Sec. IV-B"},{"comment":"The PSNR column is labeled 'PNSR'; should be PSNR.","section":"Table I"},{"comment":"The text uses MBE (maximum Bayesian estimation) while Fig. 4 uses MLE; please make the terminology consistent.","section":"Sec. V-D, Fig. 4"},{"comment":"Both tables have the caption 'Performance Comparison of Edge Sampling and Random Sampling,' but Table IV compares Hybrid (Vertex + Random) versus Random. The captions should be distinct and correct.","section":"Tables III and IV"},{"comment":"The input tensor uses two identical sparse-sampled signal strength maps; the reason for the duplication is not explained.","section":"Sec. V-B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an incremental extension of the authors' prior RadioDiff line of work. The core idea is reasonable, but the experimental protocol needs strengthening before it can support the advertised claims. In particular, the power-invariance mechanism and the comparison against classical baselines need to be reworked. The current gap between the claims and the evidence is too large for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible pipeline paper—conditional diffusion reconstructs a radio map from sparse RSS, then you read off the peak—but the two headline claims (under 1% sampling and power-invariant localization) are not supported by the experiments as reported. The power-invariance claim is especially thin: no experiment varies transmit power, and the normalization in Eq. (31) is not shown to match the training target. If the target is absolute pathloss at 23 dBm, the model can't be power-invariant; if the target is also normalized, the paper never says so or accounts for the bias from normalizing by max sampled value. That's a load-bearing gap.\n\nWhat's genuinely new: taking the RadioDiff line to a downstream localization task, with a vertex-targeted sampling heuristic and brightest-point readout. That specific pipeline is not in the prior work I know. The experiments on RadioMapSeer show the trained model can reconstruct maps and localize with a few meters error, which is nontrivial. Credit where due: the vertex sampling idea is physically motivated and the ablation in Table I does show a modest gain over random sampling at equal budget, though the gain is small (2.77 vs 3.04 m).\n\nSoft spots beyond the power issue: the main comparison in Table II is unfair. It pits a model trained on the dataset and evaluated on an upper half at ~0.96% sampling against untrained least-squares-type methods given the full map at 3.05%. That confounds method with data advantage, so the '10x fewer samples' claim isn't established. Also no learning-based baselines—RadioUNet or their own RadioDiff—which would be the natural comparison. The theoretical analysis in Sec. IV-A maximizes information about the boundary field u, not about the source position d; the link is asserted, not derived. And there are no error bars or code release, which limits reproducibility.\n\nOverall: the core idea is worth taking seriously, but the paper currently overclaims in its abstract and intro. It deserves a serious referee, and a revision that adds a varying-power experiment, normalizes or replaces the training target transparently, and redoes the baselines fairly would make the contribution solid. I'd send it to review but with a clear request for major revisions.","headline":"Plausible pipeline, but the power-invariance headline is untested and the baseline comparison is confounded.","tokens_in":16977,"tokens_out":2969,"would_cite":false,"duration_ms":32933,"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 claims that a conditional diffusion model can reconstruct a full radio map from sparse, power-normalized RSS readings taken at building corners, and that the brightest point of that map is an accurate estimate of a non-cooperative","keywords":["non-line-of-sight localization","radio map reconstruction","conditional diffusion model","sparse RSS sampling","knife-edge diffraction","power-invariant normalization","generative localization","non-cooperative emitter"],"falsifier":"Using a ray-tracing simulator that provides ground-truth emitter positions and fields, compute the Fisher information about the emitter coordinates from RSS measurements at building vertices versus at interior or random points, holding measurement noise and count fixed. If interior points are equally or more informative about the emitter position, or if random sampling at the same budget matches the localization accuracy in a controlled experiment, the central claim that vertex-targeted sampling is the source of the gain is falsified. A simpler observational check: ablate the sampling mask alo","tokens_in":16034,"feed_emoji":"📡","tokens_out":8938,"duration_ms":94447,"temperature":0.7,"pith_summary":"The paper tries to show that localizing a non-cooperative radio source in dense, non-line-of-sight environments can be solved as a generative imaging problem. Instead of estimating the emitter from individual propagation paths, it trains a conditional diffusion model to reconstruct the full radio map from a sparse set of received-signal-strength readings, then reads off the emitter as the brightest pixel. To make this work with very few measurements, it adds a physics-guided sampling rule: because diffracted energy concentrates at building edges and corners, sensors placed at those vertices carry the most information, so the required sample count scales with the scene's geometric complexity rather than its area. Unknown transmitter power is handled by normalizing all readings by the strongest observed value, making the whole pipeline power-invariant. On an urban-scale benchmark, the method reports emitter localization around three meters error at a 0.96% sampling rate, while classical RSS baselines range from 18 to 41 meters.","feed_headline":"Diffusion model finds radio sources from under 1% of samples","feed_subtitle":"Building-corner sampling plus a generative radio map recovers a hidden emitter with ~3 m error.","key_machinery":"The key machinery is the radio-map-as-prior: a conditional diffusion model that takes as input a binary building layout and two sparse RSS channels (measurements at selected locations, zeros elsewhere) and outputs a full-resolution relative signal-strength map. Sampling locations are chosen by a vertex-based mask derived from knife-edge diffraction, supported by the Fisher-information scaling J_jj proportional to 1/s_j^2 and the submodularity of mutual information, which the paper uses to argue that edge and vertex probes are information-dense anchors. Power invariance comes from normalizing each reading by the maximum observed value, and localization is the argmax of the generated map, with","core_discovery":"RadioDiff-Loc's central claim is that a conditional denoising diffusion model, trained on pairs of building layouts and radio maps, can generate a full radio map from a sparse set of vertex-targeted, max-normalized RSS measurements, and that the maximum-intensity point of that generated map is an accurate estimate of the unknown emitter position. The paper argues that this is a legitimate MAP-style inference because each radio map is treated as an implicit sample of the likelihood p(R | d, H): the diffusion model learns the prior over maps conditioned on geometry, and sparse measurements condition the reverse diffusion. The physical ingredient is knife-edge diffraction: the field near buildi","pith_inferences":["A direct testable extension would compare, in a ray-tracing simulator, the Fisher information about the emitter coordinates at vertex sites versus interior sites; if interior sites are more informative about the source position, the theoretical support for vertex sampling would need revision even if the empirical gains hold.","Because the pipeline outputs a power-invariant dense relative field, the same trained model could also serve coverage mapping, jammer localization, or channel prediction tasks without recalibration.","The brightest-point rule assumes a single dominant source; adapting the estimator to multiple emitters or extended sources would require peak detection and assignment on the generated map, a natural next step.","The sampling-efficiency claim depends on the number of vertices, so environments with many small structures may see the advantage shrink; clustering nearby vertices into a bounded set of probes would be a practical extension."],"forward_implications":["Sampling cost scales with the number of building vertices or edges rather than the number of grid cells, so simple scenes need fewer than 1% of positions measured.","No transmitter power calibration, pilot signals, or line-of-sight paths are required; only relative RSS values are needed.","The reconstructed radio map is a dense surrogate field, so classical methods such as trilateration, fingerprinting, and threshold-centroid estimators can be applied on top, forming a dual data-driven and model-driven pipeline.","The reported margin over classical baselines—about 3 meters versus 18 to 41 meters at a 0.96% sampling rate on an urban dataset—suggests that RSS-only NLoS localization is feasible under severe sensing constraints.","Under shrinking sampling budgets, the vertex-based strategy degrades more gracefully than random sampling, indicating that the sampling rule itself, not just the generative model, is what preserves accuracy."],"supporting_citations":[{"why":"Supplies the urban radio-map dataset with ground-truth pathloss used to train and evaluate the framework.","marker":"[39]"},{"why":"Supplies the ray and knife-edge diffraction theory that motivates placing sensors at building edges and vertices.","marker":"[11]"},{"why":"Provides the generative diffusion formulation for radio map construction that this paper adapts to conditional sparse-RSS localization.","marker":"[17]"},{"why":"Provides the DDPM discrete forward and reverse formulation used as the generative backbone.","marker":"[36]"},{"why":"Supplies the score-based SDE formulation and reverse-time sampling that underpin the conditional diffusion model.","marker":"[37]"},{"why":"Supplies the decoupled diffusion process, where the signal first decays to zero and then noise is added, used for stable radio-map synthesis.","marker":"[38]"},{"why":"Defines the least-squares RSS localization baseline that RadioDiff-Loc is compared against.","marker":"[40]"},{"why":"Defines the adaptive weighted least-squares baseline used for comparison.","marker":"[41]"},{"why":"Defines the maximum Bayesian estimation baseline used for comparison.","marker":"[42]"},{"why":"Defines the nonlinear least-squares baseline used for comparison.","marker":"[43]"}],"fun_headline_variants":["Diffusion model maps hidden radio sources with sparse sensing","Building-corner sampling plus diffusion pinpoints NLoS emitters","Sparse radio maps, full localization: diffusion does it","Diffusion-based RM generation cuts samples for NLoS localization","Power-invariant radio map from sparse reads localizes emitters"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The sampling strategy assumes that the places carrying the most information about the radio field near building corners are the same places carrying the most information about where the emitter is, but the paper does not prove that link for the source position itself.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model maps hidden radio sources with sparse sensing","Building-corner sampling plus diffusion pinpoints NLoS emitters","Sparse radio maps, full localization: diffusion does it","Diffusion-based RM generation cuts samples for NLoS localization","Power-invariant radio map from sparse reads localizes emitters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2678,"prompt_tokens":783,"completion_tokens":1895,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":1813}},"tokens_in":527,"tokens_out":1895,"duration_ms":15285,"temperature":1.0,"reasoning_tokens":1813,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:05:40.399879+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Using a ray-tracing simulator that provides ground-truth emitter positions and fields, compute the Fisher information about the emitter coordinates from RSS measurements at building vertices versus at interior or random points, holding measurement noise and count fixed. If interior points are equally or more informative about the emitter position, or if random sampling at the same budget matches the localization accuracy in a controlled experiment, the central claim that vertex-targeted sampling is the source of the gain is falsified. A simpler observational check: ablate the sampling mask alo","supporting_citations":[{"cited_title":"The first pathloss radio map prediction challenge,","cited_arxiv_id":null,"evidence_quote":"Supplies the urban radio-map dataset with ground-truth pathloss used to train and evaluate the framework."},{"cited_title":"Ray techniques in electromagnetics,","cited_arxiv_id":null,"evidence_quote":"Supplies the ray and knife-edge diffraction theory that motivates placing sensors at building edges and vertices."},{"cited_title":"RadioDiff: An effective generative diffusion model for sampling- free dynamic radio map construction,","cited_arxiv_id":null,"evidence_quote":"Provides the generative diffusion formulation for radio map construction that this paper adapts to conditional sparse-RSS localization."},{"cited_title":"Decoupled diffusion models: Simultaneous image to zero and zero to noise,","cited_arxiv_id":null,"evidence_quote":"Supplies the decoupled diffusion process, where the signal first decays to zero and then noise is added, used for stable radio-map synthesis."},{"cited_title":"Linear least squares localization in sensor networks,","cited_arxiv_id":null,"evidence_quote":"Defines the least-squares RSS localization baseline that RadioDiff-Loc is compared against."},{"cited_title":"Robust visual tracking using structurally random projection and weighted least squares,","cited_arxiv_id":null,"evidence_quote":"Defines the adaptive weighted least-squares baseline used for comparison."},{"cited_title":"Indoor location fingerprinting algorithm based on path loss parameter estimation and bayesian infer- ence,","cited_arxiv_id":null,"evidence_quote":"Defines the maximum Bayesian estimation baseline used for comparison."},{"cited_title":"A kernel method to nonlinear location estimation with rss-based fingerprint,","cited_arxiv_id":null,"evidence_quote":"Defines the nonlinear least-squares baseline used for comparison."}],"review_version":1}