{"id":"866f37d5-a7c8-466b-984e-da922d2ca6d1","arxiv_id":"2508.13007","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Doppler-guided sparse queries with two-stage offsets let CAVs share about 10 percent of the BEV feature data at matching detection accuracy.","lead":"SlimComm cuts cooperative driving bandwidth by sharing only sparse radar-guided queries instead of full maps, while keeping object detection accuracy. It also releases two new radar-augmented simulated datasets so other teams can test their own bandwidth-efficient collaborative perception systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 compares SlimComm (LiDAR+Doppler radar) against LiDAR-only baselines, so the claimed accuracy parity/superiority may stem from the extra sensor modality rather than the sparse-query scheme.","rationale":"The reader's concern about simulated Doppler realism is legitimate but secondary: it affects external validity, not the internal comparison. The more immediately load-bearing issue is that Table 2 may compare a multi-sensor system against LiDAR-only baselines without controlling for the extra radar modality. If that is true, the claim 'matching or surpassing prior baselines at 90% lower bandwidth' is not established even under a perfect simulator, because the accuracy component of the claim is confounded with sensor input. The paper does provide useful supporting evidence: component ablations show exploratory queries and halo enrichment contribute, robustness experiments show graceful degradation under pose noise and delay, and the datasets are a plausible contribution. Those pieces support the system's internal operation, but they do not resolve the sensor-fairness confound. The verdict should remain CONDITIONAL, with the condition expanded to require either radar-augmented baselines or a LiDAR-only SlimComm ablation. I therefore agree only partially with the reader's chosen weakest assumption: the identified realism gap is real, but the stronger threat sits in the experimental setup itself.","tokens_in":8859,"tokens_out":9248,"duration_ms":105015,"concrete_test":"Run a controlled comparison on OPV2V-R and Adver-City-R in which the four baseline methods receive the same LiDAR+radar input as SlimComm (e.g., by concatenating radar points with Doppler as extra features in the point-cloud encoder), retrained with their original settings. If the AP@0.5/AP@0.7 gaps in Table 2 shrink substantially, or if a LiDAR-only ablation of SlimComm loses its current margin (e.g., OPV2V-R AP@0.7 drops below the 0.80 reported, or the Adver-City General AP@0.7 gap of 0.54 vs 0.33 disappears), then the headline accuracy claim is not attributable to the sparse-query scheme. Report the per-query counts and bandwidth per frame alongside the results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SlimComm matches or surpasses prior baselines while using roughly 10% of their bandwidth. Table 2 reports SlimComm against AttFusion, S-AdaFusion, SCOPE, and Where2Comm. All four baselines were originally published as LiDAR point-cloud cooperative-perception methods (Refs. [26], [16], [27], [6]), and no passage in the supplied manuscript states that these baselines were modified to consume the newly added 4D radar or its Doppler channel. Meanwhile, the paper's own contributions and related work emphasize 'LiDAR–radar cooperation' and Doppler-driven query placement (Section 2.2, Table 1). If the baselines receive only LiDAR while SlimComm receives LiDAR plus Doppler-enriched radar points, the comparison is not apples-to-apples: some of the AP gains and even the apparent robustness of query placement could be due to the additional sensor modality, not to the sparse-query communication strategy. The ablations in Tables 3–4 isolate exploratory queries and halo enrichment, but they do not include a 'SlimComm without radar' condition, and no radar-augmented baseline is reported. This is a more direct threat to the internal validity of the headline claim than the simulated-radar realism issue: even if C-Shenron were a perfect radar simulator, the current evidence would not show that the sparse-query scheme alone accounts for the reported accuracy–bandwidth trade-off. The promised code and exact query-budget parameters would help reproduction, but the missing sensor-fair comparison is the load-bearing gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"SlimComm is a cooperative 3-D perception framework that uses 4D radar Doppler to construct a motion-centric dynamic map and generate two types of sparse queries: reference queries on dynamic or high-confidence regions and exploratory queries into occluded areas. Only query-specific BEV features are exchanged and fused through a gated multi-scale deformable attention module. The authors introduce OPV2V-R and Adver-City-R, CARLA-based extensions of existing V2X suites with per-point Doppler radar, and report up to 90% lower bandwidth than full-map sharing while maintaining competitive AP@0.5 and AP@0.7 against AttFusion, S-AdaFusion, SCOPE, and Where2Comm. The paper also evaluates robustness to localization noise, heading noise, and asynchronous delays up to 600 ms.","tokens_in":9191,"tokens_out":7277,"duration_ms":68398,"significance":"If validated, the Doppler-guided sparse-query idea is a useful contribution to bandwidth-efficient V2X perception, and the release of radar-augmented CARLA datasets could support further research on multi-agent 4D radar fusion. The paper provides clear ablations for the exploratory-query and halo-enrichment components, and the code and dataset release are valuable for reproducibility. However, the central empirical claim is not yet fully supported because the main comparison is confounded by sensor modality and the evaluation relies entirely on synthetic radar with no real-world validation.","major_comments":[{"comment":"The headline comparison is not apples-to-apples: SlimComm fuses LiDAR with 4D radar (with Doppler), while AttFusion, S-AdaFusion, SCOPE, and Where2Comm are LiDAR-only cooperative-perception methods. The manuscript does not state that these baselines were modified to consume the radar points or the Doppler channel. Consequently, the reported accuracy parity/superiority and the apparent robustness of query placement could stem from the additional sensor modality rather than from the sparse-query communication scheme. To support the central claim, please add a LiDAR-only variant of SlimComm as a control and, ideally, radar-augmented versions of the baselines so that all methods receive the same sensor inputs.","section":"Section 5.2.1, Table 2"},{"comment":"The ablation study in Table 3 decomposes the full model by removing exploratory queries and halo enrichment, but it never removes the radar input or the Doppler-derived dynamic map. Without a \"SlimComm without radar\" condition, the improvement over the LiDAR-only baselines cannot be attributed to the Doppler-guided query mechanism. Please add such an ablation, and if feasible an ablation with radar but without Doppler gating, to isolate the contribution of the proposed sparse-query scheme from the contribution of the extra sensor modality.","section":"Section 5.3, Table 3"},{"comment":"All experiments are conducted on CARLA-based datasets with radar data generated by the C-Shenron simulator, and the manuscript does not provide any validation of the simulated per-point Doppler against real 4D radar measurements. Since the method's query placement and motion-centric filtering rely on the Doppler channel, the transferability of the reported results to real radar is unverified. Please state this limitation explicitly and, if possible, include a small-scale real-radar experiment or an analysis of simulator-to-real gaps.","section":"Section 3 and Table 1"},{"comment":"The communication cost is reported as \"CV (measured in log2 scale)\" and \"BD (in MB/frame)\", but the text does not define the exact computation: per-query payload size, halo dimensions, feature bit depth or quantization, number of collaborating agents, or averaging over frames. Without this protocol, the central quantitative claim of \"up to 90% lower bandwidth\" cannot be independently reproduced. Please provide the precise bandwidth formula and the query-budget, halo, and offset parameters used for the results.","section":"Section 5.2.1, Table 2"}],"minor_comments":[{"comment":"On OPV2V-R, SlimComm's AP@0.5 is 0.87 versus S-AdaFusion's 0.91, so the claim of \"matching or surpassing prior baselines\" is slightly overstated; please rephrase to \"comparable or better\" or report statistical significance.","section":"Abstract and Table 2"},{"comment":"The Adver-City-R columns mix three scenario-specific values for AP and only two scalar values for CV and BD; the presentation makes it difficult to determine which CV/BD value corresponds to which scenario. Please use separate rows or columns.","section":"Table 2"},{"comment":"The text states \"less than 1.5 mAP degradation\" while the figure and metric are AP; please keep the metric name consistent.","section":"Section 5.2.2"},{"comment":"The C-Shenron simulator is cited, but no specific simulation parameters are given for the radar noise model or Doppler computation; please include these details so that the dataset generation is reproducible.","section":"Section 3 and Supplementary Material"}],"recommendation":"major_revision","confidential_remarks":"The central claim is currently confounded by the sensor-modality mismatch in Table 2. Even if the method is promising, the manuscript should not be accepted until this is resolved with matched-sensor experiments. The released datasets may be useful to the community, but the synthetic-radar caveat should be communicated clearly in the final version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core idea is genuinely new: using Doppler radar to place sparse queries in cooperative perception, plus an occlusion-shadow branch that probes behind occluders. That's a real step beyond the confidence/entropy heuristics in Where2Comm and EffiComm. Second, the headline comparison in Table 2 is apples-to-oranges. SlimComm is trained with LiDAR plus Doppler radar; the baselines (AttFusion, S-AdaFusion, SCOPE, Where2Comm) are LiDAR-only methods with no radar channel. No passage says the baselines were adapted to consume radar, and there is no 'SlimComm without radar' ablation. So the matching-or-better accuracy at ~10% bandwidth may reflect the extra sensor modality, not the sparse-query scheme. That is the load-bearing gap.\n\nThe paper also does some things well. The ablations are honest: they show the exploratory branch and halo enrichment each add a few AP points, and they report bandwidth in both log scale and MB/frame. The datasets, OPV2V-R and Adver-City-R, are a real contribution—CARLA-based with per-point Doppler, and the paper is upfront that they are simulated. The robustness experiments under localization noise and async delays are useful, though the delay curves are relegated to the supplement.\n\nThe soft spots are three, in decreasing importance. First, the sensor-fairness problem above. If the method is meant to be a full LiDAR+radar system, then compare it against LiDAR+radar baselines, or at least ablate the radar. Second, simulated radar realism: C-Shenron is a plausible simulator, but no evidence is given that its Doppler behavior transfers to real 4D radar, so the motion-prior benefits could be simulation artifacts. Third, the exact query budget and offset parameters are missing, and the code/data are promised but not released. That's minor if they do release, but it makes reproduction currently impossible.\n\nOverall: this is a solid systems paper with an interesting idea and useful datasets, but the central empirical claim is not yet supported. It deserves a serious referee—the novelty is real and the datasets have standalone value—but the revision should require a radar-less SlimComm variant and at least one radar-augmented baseline. I'd send it to review, not desk reject, and tell the authors exactly what's missing.","headline":"The accuracy-bandwidth claim is undercut by an unfair sensor comparison, but the Doppler-guided query idea and the two new datasets are genuine contributions worth peer review.","tokens_in":9729,"tokens_out":3114,"would_cite":true,"duration_ms":30300,"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":"SlimComm's Doppler-guided sparse query scheme shares only small feature patches, cutting cooperative-perception bandwidth by up to 90 percent while matching full-map detection accuracy.","keywords":["collaborative perception","4D radar","Doppler","sparse queries","bandwidth efficiency","BEV feature communication","occlusion-aware perception","3D object detection"],"falsifier":"Equip a test fleet with per-point Doppler radar and run SlimComm on recorded scenes with dense moving and static traffic; if the motion map routes queries to static clutter or misses moving objects, the reported accuracy at low bandwidth will not reproduce.","tokens_in":8684,"feed_emoji":"🚗","tokens_out":6972,"duration_ms":66882,"temperature":0.7,"pith_summary":"SlimComm aims to show that connected vehicles can cooperate over limited V2X bandwidth without abandoning accuracy: instead of exchanging dense bird's-eye-view feature maps, each agent broadcasts a small set of queries, and only the requested feature patches are returned. The paper's central proposal is to let 4D radar Doppler decide where queries go—its ego-motion-compensated signal separates moving from static objects, while a confidence prior plants extra queries in occlusion shadows. The authors release two simulated benchmarks, OPV2V-R and Adver-City-R, with per-point Doppler radar, and report that SlimComm consumes up to 90 percent less bandwidth than full-map sharing while matching or exceeding prior cooperative perception methods in average precision. If the results hold, bandwidth-efficient cooperative perception becomes a query-selection problem rather than a compression problem, with motion and occlusion semantics doing the selection.","feed_headline":"Sparse Doppler-guided queries cut cooperative-vision bandwidth 90%","feed_subtitle":"Connected cars share only small feature patches around moving and hidden objects, keeping detection accuracy.","key_machinery":"The load-bearing mechanism is the Doppler-guided query generator. Ego-motion compensation turns radar returns into a binary motion prior: moving objects carry a nonzero Doppler residual, static background does not, so the map is motion-centric rather than confidence-only. Two query types are produced: reference queries selected on dynamic or high-confidence cells, and exploratory queries sampled in occlusion shadows cast by strong confidence peaks. A two-stage offset—a coarse nudge followed by a 3×3 deformable halo—turns each anchor into a compact sampling cluster, and the gated multi-scale deformable attention block fuses the returned patches. The mechanism's job is to concentrate the exchanged bits where evidence is strongest or likely hidden, leaving static background unsampled.","core_discovery":"The central discovery is that motion and occlusion semantics, read off per-point radar Doppler, can drive a sparse communication protocol that keeps dense-map accuracy. SlimComm builds a motion-centric dynamic map from ego-motion-compensated Doppler returns, places reference queries on dynamic cells and high-confidence foreground cells, and places exploratory queries behind confidence peaks so collaborators return context for partially or fully occluded regions. Neighbours warping their features into the ego frame send only a 3×3 halo-enriched patch per query, and a multi-scale gated deformable attention block fuses the responses. On the released OPV2V-R and Adver-City-R datasets the scheme reports up to 90% bandwidth reduction relative to full-map sharing, with AP@0.5 and AP@0.7 matching or exceeding strong baselines, and degrades by less than 1.5 mAP under 200–600 ms asynchronous delays.","pith_inferences":["Editorial inference: the same query logic should transfer to setups without radar by substituting optical flow or learned motion cues, since the essential signal is a cheap motion-versus-static prior rather than radar specifically.","Editorial inference: on real roads with many stationary objects, the bandwidth saving should grow with traffic density, because queries concentrate on movers and occlusion shadows rather than background; a real-radar benchmark could test this.","Editorial inference: the query budget itself could be tuned dynamically from scene statistics, for example raising the exploratory-query count at intersections where occluders are dense."],"forward_implications":["Bandwidth per frame drops to roughly 10% of full-map sharing: on OPV2V-R, SlimComm uses 0.63 MB/frame versus 4.45 MB for the next most efficient baseline and tens of MB for dense-map methods.","Accuracy is not traded away: on Adver-City-R SlimComm reaches AP@0.5 of 0.67 and AP@0.7 of 0.54 in the general split, at or above AttFusion and S-AdaFusion, and on OPV2V-R it stays competitive while using 7x less bandwidth than Where2Comm.","Exploratory queries materially help occluded cases: ablations show the exploratory branch lifts AP@0.7 on Adver-City-R from 0.41 to 0.49, and the halo enrichment lifts it further to 0.54.","Stability under pose and timing errors: SlimComm maintains the best AP@0.7 across localization noise up to 0.6 m and heading noise up to 1.0°, and loses less than 1.5 mAP under delays up to 600 ms."],"supporting_citations":[{"why":"Supplies the OPV2V base dataset and the AttFusion full-map sharing baseline that SlimComm extends with radar and must match in accuracy.","marker":"[26]"},{"why":"Supplies the Adver-City base dataset used to build the harder Adver-City-R benchmark.","marker":"[8]"},{"why":"Provides the radar simulation model used to add per-point Doppler returns to the simulated scenes.","marker":"[14]"},{"why":"Provides the open urban driving simulator in which OPV2V-R and Adver-City-R are generated.","marker":"[4]"},{"why":"Provides the cooperative driving simulation pipeline that synchronizes agents and sensors for dataset creation.","marker":"[25]"},{"why":"Defines the Where2Comm sparse-confidence baseline that SlimComm compares against and outperforms in the bandwidth-accuracy trade-off.","marker":"[6]"}],"fun_headline_variants":["Doppler-guided sparse queries cut cooperative vision bandwidth 90%","Radar Doppler steers sparse sharing in connected-car perception","SlimComm: 90% less bandwidth, same cooperative 3D vision accuracy","Sparse Doppler-guided queries keep accuracy while cutting bandwidth 90%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The simulated per-point radar Doppler reproduces real 4D radar behavior—velocity on moving objects, near-zero Doppler on static ones—so the query placement learned in simulation will transfer to physical vehicles.","fun_headline_variants_meta":{"raw":{"variants":["Doppler-guided sparse queries cut cooperative vision bandwidth 90%","Radar Doppler steers sparse sharing in connected-car perception","SlimComm: 90% less bandwidth, same cooperative 3D vision accuracy","Sparse Doppler-guided queries keep accuracy while cutting bandwidth 90%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000567,"raw_usage":{"total_tokens":2674,"prompt_tokens":922,"completion_tokens":1752,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1676}},"tokens_in":538,"tokens_out":1752,"duration_ms":13515,"temperature":1.0,"reasoning_tokens":1676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:14:57.491970+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Equip a test fleet with per-point Doppler radar and run SlimComm on recorded scenes with dense moving and static traffic; if the motion map routes queries to static clutter or misses moving objects, the reported accuracy at low bandwidth will not reproduce.","supporting_citations":[{"cited_title":"Opv2v: An open benchmark dataset and fusion pipelineforperceptionwithvehicle-to-vehiclecommunication","cited_arxiv_id":null,"evidence_quote":"Supplies the OPV2V base dataset and the AttFusion full-map sharing baseline that SlimComm extends with radar and must match in accuracy."},{"cited_title":"DemoAbstract: C-Shenron: ARealistic 11 Radar Simulation Framework for CARLA, page 726–727","cited_arxiv_id":null,"evidence_quote":"Provides the radar simulation model used to add per-point Doppler returns to the simulated scenes."},{"cited_title":"CARLA: An open urban driving simulator","cited_arxiv_id":null,"evidence_quote":"Provides the open urban driving simulator in which OPV2V-R and Adver-City-R are generated."},{"cited_title":"Opencda: an open cooperative driving automation framework integrated with co-simulation","cited_arxiv_id":null,"evidence_quote":"Provides the cooperative driving simulation pipeline that synchronizes agents and sensors for dataset creation."},{"cited_title":"Where2comm: Communication-efficient collaborative perceptionviaspatialconfidencemaps","cited_arxiv_id":null,"evidence_quote":"Defines the Where2Comm sparse-confidence baseline that SlimComm compares against and outperforms in the bandwidth-accuracy trade-off."}],"review_version":2}