{"id":"a03d2a1a-e1da-477f-b82c-6aae18e9ba02","arxiv_id":"2606.11914","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"NARRAS is a learned edge-triggered reporting policy for distributed CSI-based localization that achieves better accuracy than baselines at the same uplink activity level using recurrent state and channel chart regularization.","lead":"The paper proposes NARRAS, a decentralized policy where each remote antenna array uses recurrent summaries and memory of past transmissions to decide whether to report CSI features for vehicle position estimation under an activity budget. Smart generalists might read it to see how edge devices can coordinate inference tasks without flooding shared communication channels.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"Reader correctly notes that only the abstract was available and therefore withholds a verdict. Because no concrete technical flaw in the argument can be located from the supplied material, the UNVERDICTED status is left unchanged.","tokens_in":1781,"tokens_out":245,"duration_ms":17189,"concrete_test":"Re-run the NARRAS training and evaluation pipeline on the same channel realizations used in the paper, confirming that the realized average activity exactly matches the target budget and that the reported localization error percentiles are reproduced within sampling error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract describes a coherent ETDI formulation instantiated via recurrent local summaries plus last-latent memory, trained end-to-end with differentiable activity penalties and chart-regularized latents, then thresholded at validation time. The central experimental claim (improved accuracy at matched uplink activity, plus benefit of chart regularization in low-activity regimes) follows directly from that construction if the reported simulations are executed as described. No internal inconsistency, hidden assumption in the loss, or mismatch between the stated method and the claimed outcome is visible in the given description.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces the Edge-Triggered Distributed Inference (ETDI) abstraction for CSI-based UE localization with distributed remote antenna arrays (RAAs) under uplink activity budgets. It proposes NARRAS, a decentralized policy that combines recurrent summaries of local observations with memory of the last transmitted latent, trained end-to-end via differentiable activity penalties, validation-calibrated thresholds, and channel-chart regularization of the latent space. The central claim is that NARRAS achieves higher localization accuracy than learned and heuristic sparse-reporting baselines at matched uplink activity, with chart regularization further reducing high-percentile errors in low-activity regimes.","tokens_in":1894,"tokens_out":394,"duration_ms":12062,"significance":"If the empirical claims are substantiated with reproducible experiments, the work would contribute a concrete instantiation of task-oriented communication for vehicular IoT localization, demonstrating the value of local recurrent decision rules and geometry-aware latents under communication constraints. The explicit activity-budget control via penalties and the separation of training from validation-time thresholding are methodologically clean elements.","major_comments":[{"comment":"Abstract: the central empirical claim (improved accuracy at comparable uplink activity, plus benefit of chart regularization in low-activity regimes) is stated without any quantitative results, error bars, dataset descriptions, baseline specifications, or experimental-setup details. This absence is load-bearing because the paper's contribution rests on the reported performance gains over baselines.","section":"Abstract"},{"comment":"The weakest assumption—that the recurrent local decision rule (recurrent summary + last latent) trained with activity penalties and calibrated thresholds will generalize to unseen channel conditions without central coordination—is not stress-tested in the provided description. A concrete cross-scenario or out-of-distribution evaluation would be needed to support the claim that the policy remains effective in real deployments.","section":"Abstract"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed review and constructive feedback on our manuscript. We address the major comments below and outline the revisions we plan to make.","responses":[{"response":"We agree that the abstract would benefit from including quantitative results to substantiate the claims. We will revise the abstract to incorporate specific performance metrics, such as localization accuracy improvements at given activity levels, along with brief mentions of the dataset and baselines used.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central empirical claim (improved accuracy at comparable uplink activity, plus benefit of chart regularization in low-activity regimes) is stated without any quantitative results, error bars, dataset descriptions, baseline specifications, or experimental-setup details. This absence is load-bearing because the paper's contribution rests on the reported performance gains over baselines."},{"response":"Our evaluations are conducted on the scenarios described in the paper. We acknowledge that explicit out-of-distribution testing for unseen channel conditions is not presented. We will include a note in the discussion section of the revised manuscript highlighting this as a direction for future validation, while noting that the training methodology with recurrent summaries aims to support generalization.","revision_made":"partial","referee_comment":"[Abstract] The weakest assumption—that the recurrent local decision rule (recurrent summary + last latent) trained with activity penalties and calibrated thresholds will generalize to unseen channel conditions without central coordination—is not stress-tested in the provided description. A concrete cross-scenario or out-of-distribution evaluation would be needed to support the claim that the policy remains effective in real deployments."}],"tokens_in":1460,"tokens_out":350,"duration_ms":18341,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The key takeaway is that NARRAS combines recurrent state with memory of the last transmitted latent to make local decisions on whether to report CSI features, trained end-to-end with activity penalties and channel chart regularization. This seems to give better accuracy than other sparse strategies at the same activity level, and the regularization helps in very low activity cases.\n\nWhat the paper does well is laying out the ETDI abstraction clearly and showing how to train the policy without central coordination while respecting the budget. The use of differentiable penalties and validation thresholds is a practical way to handle the constraint. The channel chart regularization is a nice touch to make the latents more robust when few are reported.\n\nThe soft spots are minor but worth noting. The abstract mentions improvements but doesn't give the actual error values or the exact baselines, so the magnitude of the gain is not clear from the summary. Also, the experiments are simulations, and the claim about robustness in low-activity regimes would benefit from more details on how the chart is built and whether it transfers to new environments. The assumption that the local rule generalizes without retraining is plausible but not proven in the description.\n\nThis is the kind of paper that fits in wireless sensing and edge computing venues. Readers working on distributed inference or resource allocation in IoT will find the formulation useful. It deserves a serious referee because the method is well-motivated and the approach is technically sound on its own terms.\n\nI would send it out for review.","headline":"NARRAS gives a recurrent local policy for edge-triggered CSI reporting that improves accuracy at matched uplink activity, with chart regularization helping in sparse regimes.","tokens_in":2366,"tokens_out":373,"would_cite":false,"duration_ms":15431,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Each remote antenna array decides locally whether to report its CSI observation, improving localization accuracy at a fixed uplink activity budget.","keywords":["CSI-based localization","distributed inference","edge-triggered reporting","vehicular IoT","channel charting","sparse reporting","NARRAS","task-oriented communication"],"falsifier":"A direct accuracy comparison on a new vehicular dataset collected under different mobility patterns or propagation conditions would show whether the reported gains over baseline sparse strategies persist.","tokens_in":2686,"feed_emoji":"📡","tokens_out":681,"duration_ms":17466,"temperature":0.7,"pith_summary":"The paper sets out to show that CSI-based localization in vehicular IoT networks can be performed more efficiently by letting each distributed antenna array make its own decision on whether to forward its current observation. This is framed as Edge-Triggered Distributed Inference, an abstraction for any task where resource-limited devices share an access channel for a joint inference goal. The concrete policy, called NARRAS, combines a recurrent summary of recent local observations with memory of the last latent feature it sent, and is trained under an explicit average-activity constraint plus channel-chart regularization. A sympathetic reader would care because the shared uplink can support only limited simultaneous transmissions, so any method that cuts unnecessary reports without sacrificing accuracy directly relaxes a fundamental resource bottleneck.","feed_headline":"Decentralized CSI reporting improves vehicle positioning at fixed uplink activity","feed_subtitle":"NARRAS lets each antenna array trigger reports using local recurrent summaries and latent memory, outperforming other sparse methods especia","key_machinery":"NARRAS, the decentralized reporting policy that uses a recurrent summary of observations combined with memory of the last latent transmitted, controlled by activity penalties and channel-chart regularization.","core_discovery":"NARRAS instantiates edge-triggered distributed inference for CSI localization by having each RAA combine a recurrent summary of recent observations with memory of its last transmitted latent feature. Training uses differentiable activity penalties and validation-calibrated thresholds to enforce the budget, along with channel-chart regularization for latent geometry. This yields improved position estimation accuracy compared to learned and heuristic sparse-reporting methods at comparable uplink activity, with regularization further reducing high-percentile errors in low-activity regimes.","pith_inferences":["The same local-trigger structure could be applied to other distributed sensing tasks that share a contention-based uplink.","Geometry-aware latent representations shaped by channel charting may remain useful even when reporting becomes sparser than the training regime.","Removing the need for central coordination at every time step could allow the scheme to scale to larger numbers of remote antenna arrays."],"forward_implications":["At comparable uplink activity, NARRAS improves localization accuracy over learned and heuristic sparse-reporting strategies.","In low-activity regimes, chart regularization reduces high-percentile localization errors.","Dense full-report models remain useful as budget-free performance references.","The ETDI abstraction applies to a broader class of task-oriented communication problems."],"fun_headline_variants":["NARRAS triggers CSI reports via recurrent summaries and latent memory","Edge-triggered inference controls CSI activity for vehicle positioning","Recurrent RAA memory enables budgeted sparse CSI localization","Channel-chart regularization shapes latents under activity constraints"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The local decision rule based on recurrent summary and last latent, trained with activity penalties and calibrated thresholds, will generalize to unseen channel conditions and real-world deployments without central coordination.","fun_headline_variants_meta":{"raw":{"variants":["NARRAS triggers CSI reports via recurrent summaries and latent memory","Edge-triggered inference controls CSI activity for vehicle positioning","Recurrent RAA memory enables budgeted sparse CSI localization","Channel-chart regularization shapes latents under activity constraints"]},"model":"grok-4.3","cost_usd":0.005122,"raw_usage":{"total_tokens":2536,"prompt_tokens":759,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":51224500,"prompt_tokens_details":{"text_tokens":759,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1716,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":759,"tokens_out":61,"duration_ms":10431,"temperature":1.0,"reasoning_tokens":1716,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T08:51:50.184335+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A direct accuracy comparison on a new vehicular dataset collected under different mobility patterns or propagation conditions would show whether the reported gains over baseline sparse strategies persist.","supporting_citations":[],"review_version":1}