{"id":"ea56a6c2-f229-410a-a5d1-baa1e2906adb","arxiv_id":"2507.19354","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"EffiComm prunes BEV feature maps using a confidence mask and a graph-based adaptive keep-ratio policy, fuses survivors with Mixture-of-Experts attention, and reports 0.843 AP@0.7 on OPV2V while sending about 1.90 MB per frame.","lead":"This paper presents EffiComm, a system that lets connected vehicles share 3D sensor maps using far fewer bytes by pruning uninformative regions before transmission. It matters because bandwidth limits how well cars can cooperate in real traffic, and EffiComm reports a large reduction in transmitted data while keeping detection accuracy roughly unchanged.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline '<40%' rests on an undefined MB metric; abstract's 1.5 MB conflicts with Table III's 1.90 MB, so a wire-format byte count is needed before the central claim can be accepted.","rationale":"I agree with the reader that the load-bearing premise is the fidelity of the bandwidth measurement. The abstract's quantitative claim depends entirely on comparing MB-per-frame across methods, but the paper specifies no method for producing Table III's MB values. The definition in Section IV-A is about log2 of non-zero counts, and Table IV's '×B' is undefined; the numbers are internally coherent only if B≈1000 bytes/cell, which is a dense feature-byte proxy, not a radio-level byte count. Independent support in the paper is limited: no released code, no machine-checked proofs, no third-party evaluation; the OPV2V numbers are taken at face value. The concern is addressable by a concrete serialization experiment, so the suitable verdict is the same CONDITIONAL the reader gave. I do not see grounds for REJECT: the method's accuracy results and the relative ordering of bandwidth are plausible and reproducible in principle, and the missing measurement definition is fixable without changing the architecture. The abstract's 1.5 MB vs Table III's 1.90 MB is an internal inconsistency that should be corrected as part of the revision.","tokens_in":10847,"tokens_out":10488,"duration_ms":106996,"concrete_test":"Run a wire-format simulation on OPV2V test frames with the released EffiComm and Where2Comm checkpoints: serialize the actually transmitted F'_v tensors with a concrete format (e.g., 16-bit values plus run-length/bitmask indices and per-message headers) and count end-to-end bytes per frame; recompute the EffiComm/Where2Comm ratio. Also instrument the code to reproduce the abstract's 1.5 MB figure from any table row; if it cannot be reproduced, the abstract must be corrected to 1.90 MB (or 1.37 MB for the +AL variant).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section IV-A defines the primary communication metric as log2 of non-zero element count, 'simulating realistic data compression effects,' but the MB numbers in Table III, the actual evidence for the '<40% of prior art' claim, are never tied to a defined measurement procedure. The only formula clue, Table IV's footnote 'log2(non-zero elements×B)', leaves B undefined; solving the rows against Table III implies B ≈ 1000 bytes per retained cell (256 channels × 4 bytes), i.e., a raw dense-tensor byte count for surviving cells, not a V2V wire format. No allowance is made for coordinate indices or bitmasks, quantization or packetization, or the extra signaling needed so that all vehicles can run the GAT-based AGR before transmission. The abstract's 'approximately 1.5 MB per frame' also matches no entry in Table III (1.90 MB) or Table V (1.37 MB for the +AL variant), so the headline number is not reproducible from the tables. If the intended claim is about airtime, the measurement must be specified; if it is only a raw feature-byte proxy, the paper should say so and qualify '<40%' accordingly.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EffiComm, an end-to-end framework for bandwidth-efficient V2V collaborative perception. It operates on BEV feature maps, applies a two-stage reduction pipeline (Selective Transmission, which prunes low-confidence spatial cells, and Adaptive Grid Reduction, which uses a graph attention network to assign per-vehicle keep ratios conditioned on role and network load), and fuses the surviving features with a soft-gated Mixture-of-Experts attention layer. On the OPV2V test set, the authors report AP@0.7 of 0.843 with a mean communication cost of 1.90 MB per frame, compared with 0.83 and 4.9 MB for Where2Comm, and claim that EffiComm transmits less than 40% of the data required by prior art while maintaining state-of-the-art detection accuracy. The paper also reports results on the Culver City dataset, an ablation of the MoE fusion head, and a variant trained with additional bandwidth and entropy losses.","tokens_in":11132,"tokens_out":5243,"duration_ms":55380,"significance":"If the communication-cost measurement were transparent and reproducible, this would be a solid contribution to collaborative perception: it shows that end-to-end training with only a detection loss can produce emergent sparsity that reduces transmitted features by roughly a factor of 2.5 relative to Where2Comm with a small accuracy gain, and the use of a GNN to output per-vehicle keep ratios is a plausible mechanism for load-aware compression. The evaluation is held-out on a standard benchmark, and the promised public release of code and weights would aid reproducibility. However, the central claim of the paper is a bandwidth claim, and the current manuscript does not define the byte-level measurement that supports it; the headline numbers in the abstract and the tables are not mutually consistent. The significance of the work can only be assessed after this measurement issue is resolved.","major_comments":[{"comment":"The central efficiency claim, that EffiComm 'transmits less than 40% of the data required by prior art,' is supported by the MB column in Table III, but no measurement procedure for these MB values is specified. Section IV-A defines the primary communication metric as log2 of the number of non-zero elements, which is not a byte count, and then states that raw bandwidth is 'directly measured' without describing the measurement: is it the number of retained feature elements times bytes per element, the size of a serialized sparse tensor, or an actual over-the-air V2V payload? No accounting is given for index coordinates, bitmasks, quantization, packetization, or the signaling needed for vehicles to know each other's masks and keep ratios. Because the headline ratio 1.90 MB vs 4.9 MB depends entirely on this definition, please specify the wire format or, failing that, an exact reproducible tensor-byte counting rule, and recompute the ratio under it. Also reconcile the abstract's 'approximately 1.5 MB per frame' with Table III's 1.90 MB and Table V's 1.37 MB; as written, the headline number matches no table entry.","section":"§IV-A, Tables III-V"},{"comment":"The measurement definition is internally inconsistent. Table IV's footnote defines Comm as log2(non-zero elements × B) but leaves B undefined, while Tables I and II report Comm as a bare log2 count without any B factor. If B is bytes per element, then the log2 values in Table I are not directly comparable to the MB measurements in Table III, because each retained grid cell is a multi-channel feature vector rather than a single bit; if B is something else, it must be stated. As written, the reader cannot convert the reported log2 costs into MB, cannot check the claimed accuracy-per-bit trade-off, and cannot verify that the comparison to Where2Comm uses the same cost definition for all methods. Please use a single, clearly defined cost metric throughout, or report both the raw sparsity proxy and the byte-level measurement separately with their exact formulas.","section":"Table IV footnote, Table I"},{"comment":"The comparison with prior methods may not be on equal footing because the manuscript does not state whether the same MB measurement procedure was applied to all baselines. Section IV-A says open-source weights were used for AttentiveFusion and FCooper and that Where2Comm was retrained, but it does not say whether the reported MB values for those methods include the same mask signaling, index overhead, or feature-channel counts as EffiComm. If the baselines' numbers were computed with different feature-map dimensions or without their own mask overhead, the '<40% of prior art' claim would be an artifact of the comparison protocol. Please specify the exact measurement pipeline for each method, including any compression or serialization steps.","section":"§IV-A, Tables I-III"}],"minor_comments":[{"comment":"The sentence introducing node features says 'n_v = [e_hat_v, f_v, τ] where f_v is a binary flag and r represents the overall transmission rate,' but the variable r is introduced without definition and τ is not defined in that paragraph; please align the notation with Eq. (4).","section":"§III-C"},{"comment":"The text states that EffiComm's maximum observed bandwidth is 2.66 MB and that Where2Comm's standard deviation is 0.35 MB, but Table III lists 2.44 MB and 0.37 MB respectively; please correct the text or the table.","section":"§IV-D, Table III"},{"comment":"Algorithm 1 line 7 says 'for each batch b do,' which is semantically wrong for a per-frame inference procedure; this should be per scene or per frame, and the variables F_b, M_b, and τ_b should be defined with respect to vehicles and time steps.","section":"Algorithm 1"},{"comment":"The table formatting is inconsistent: on Culver City, EffiComm does not achieve the best AP, yet the text says it is 'competitive'; this is acceptable, but the tables should be explicit about which entries are best and the text should be precise about where EffiComm is not state-of-the-art.","section":"Tables I-II"},{"comment":"Reference [5] has a malformed author list ('J. L. J. M. Runsheng Xu Hao Xiang, Xin Xia Xu Han'); please correct it to the OPV2V author list.","section":"References"},{"comment":"Figure 4's caption does not define the colors used for the three keep-ratio distributions, and Figure 5's six panels are not labeled in the caption; please add legends and panel labels so the reader can follow the discussion in Sections IV-E and IV-F.","section":"Figures 4-5"}],"recommendation":"major_revision","confidential_remarks":"The core problem is that the paper's headline contribution is an efficiency claim, and the efficiency metric is undefined in a way that the authors themselves demonstrate: the abstract's 1.5 MB per frame is inconsistent with the 1.90 MB in Table III and the 1.37 MB in Table V, and the only formula clue, log2(non-zero elements × B) with B undefined, does not fix the gap. This is fixable within the scope of the manuscript by adding a precise byte-accounting procedure and recomputing all tables, but without that fix the paper should not be accepted. The detection-accuracy story is plausible and the emergent-sparsity result from detection-only training is interesting, so I would not recommend rejection if the authors can supply a transparent wire-format or tensor-byte count."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: EffiComm’s main idea is solid—replace Where2Comm’s fixed spatial threshold with a GNN that adjusts each vehicle’s keep ratio based on role and instantaneous link load, then fuse with an MoE attention head. The OPV2V numbers, taken at face value, show a nice accuracy-per-bit improvement over Where2Comm (0.843 vs 0.83 AP@0.7, 1.90 vs 4.9 MB/frame). That would be a useful step forward for bandwidth-efficient cooperative perception.\n\nWhat’s genuinely new: the learned per-vehicle keep ratios from a GAT, conditioned on the transmission rate, is a direct extension of Where2Comm’s blanket threshold and does not appear in the cited prior art. The MoE fusion is not in itself new, but applying it to sparsified multi-agent BEV features and showing it improves efficiency via sharper upstream filtering is a reasonable contribution. They also report std/max/min bandwidth, which is more than most.\n\nNow the soft spots. The bandwidth measurement is the load-bearing problem. The paper defines the primary Comm metric as log2 of non-zero element count, but the MB column in Table III has no defined measurement procedure. Table IV’s footnote mentions log2(non-zero elements×B) with B unspecified; back-solving implies B≈1000 bytes per retained cell, i.e., raw dense float32 tensors for surviving cells. There is no allowance for coordinate indices, bitmasks, quantization, packetization, or the signaling needed to run the GAT. The latter is a real issue: the AGR requires each vehicle to share node embeddings and the current τ with all other vehicles to compute keep ratios, and that coordination traffic is not counted anywhere. The abstract’s “approximately 1.5 MB per frame” also matches no number in Table III (1.90) or Table V (1.37), so the headline is not reproducible from the paper’s own tables. These are fixable: provide a wire-format byte estimate, either from a real encoder or a clear model of overhead, and reconcile the abstract.\n\nA smaller complaint: no error bars or multiple seeds, and the baseline set is thin (no V2X-ViT or DiscoGraph in the comparison, despite being cited in related work). The Where2Comm model was retrained independently, which is fine, but a release of code and weights would help.\n\nWho is this for? Researchers working on communication-efficient V2X perception. It deserves a serious referee, but the referee should push for a proper byte-level accounting and the promised code. If the metric gets fixed and the numbers hold, this is a worthwhile paper.\n\nRecommendation: send it to peer review with a request for a defined wire-format measurement and a revision that clears up the abstract discrepancy.","headline":"Promising learned per-vehicle compression for V2X, but the headline numbers rest on a fuzzy byte metric and a GAT that needs its own signaling budget.","tokens_in":11639,"tokens_out":5641,"would_cite":false,"duration_ms":49220,"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":"EffiComm claims two-stage pruning transmits less than 40% of prior art's data while keeping top 3D detection accuracy.","keywords":["collaborative perception","V2X communication","BEV feature compression","selective transmission","adaptive grid reduction","mixture of experts","graph attention network","3D object detection"],"falsifier":"Instrument a full sparse-encoding version of EffiComm—index lists for kept cells, quantized feature values, per-vehicle keep ratios, and mask signaling—and transmit over a V2V channel emulator on the OPV2V test set. If the measured mean bytes per frame is not below 40% of a similarly instrumented Where2Comm run while keeping AP@0.7 within the reported margin, the paper's central efficiency claim is wrong.","tokens_in":10652,"feed_emoji":"📡","tokens_out":6171,"duration_ms":53242,"temperature":0.7,"pith_summary":"EffiComm is an end-to-end pipeline for collaborative perception that claims to transmit less than 40% of the data required by earlier feature-sharing methods while matching or exceeding their 3D object-detection accuracy on the OPV2V benchmark. It works directly on bird's-eye-view feature maps, first pruning low-confidence spatial cells with a mask, then using a graph attention network to assign each vehicle a keep ratio that depends on its role and current network load. Surviving features are fused on the ego vehicle by a soft-gated mixture-of-experts attention layer. On OPV2V, the paper reports 0.843 AP@0.7 at roughly 1.9 MB per frame, against 0.83 at 4.9 MB for the nearest prior method. The claim matters because V2V bandwidth is the bottleneck that keeps collaborative perception from scaling to dense traffic.","feed_headline":"Two-stage pruning sends under 40% of prior data at top 3D accuracy","feed_subtitle":"On OPV2V it hits 0.843 AP@0.7 with ~1.9 MB per frame, versus 4.9 MB for the nearest spatial-masking baseline.","key_machinery":"The load-bearing mechanism is a two-stage spatial-pruning pipeline on BEV features, followed by a mixture-of-experts attention fusion. Selective Transmission computes a class-confidence importance map and keeps only cells above a fixed threshold (inference) or a random top-$K$ subset (training). Adaptive Grid Reduction then runs a graph attention network over the vehicles, taking each vehicle's confidence-derived embedding, vehicle type, and the instantaneous transmission rate, to predict a keep ratio that is used to retain the top-$K_v$ highest-confidence cells. The resulting sparse maps are fused by a softmax-gated sum of several scaled dot-product attention experts. The argument is that the first stage removes useless background, the second coordinates competing vehicles, and the MoE head recovers capacity lost to pruning.","core_discovery":"The central claim is that adaptive learned communication, not stronger sensors or bigger models, can close most of the accuracy gap between full-bandwidth feature sharing and bandwidth-limited sharing. EffiComm's two-stage reduction chooses which grid cells to keep, then how many each vehicle may keep, so that what is aired is concentrated where the detection head is most uncertain. The paper argues that the small accuracy cost of this sparsification is paid back by a soft-gated mixture-of-experts fusion head, which routes different spatial patterns to specialized attention experts and thereby integrates reduced features better than dense attention does. Empirically the paper reports the best accuracy-per-bit trade-off it compares against on OPV2V and competitive accuracy on the Culver City split, with communication cost measured as the $\\log_2$ of the number of non-zero transmitted feature elements.","pith_inferences":["The headline '<40% of prior art' rests on a specific measurement convention ($\\log_2$ of non-zero element count); a deployment planner should re-check the claim once sparse-index coding, quantization, and per-vehicle mask overhead are added to the byte count.","Because the MoE variant achieves lower bandwidth even with a higher average AGR keep ratio, the paper suggests the fusion head acts upstream as a feature regularizer; an ablation that separates the router's influence on the confidence map from its fusion capacity would isolate that effect.","The same keep-ratio feedback loop could be extended to include latency or packet-loss metrics, not just instantaneous rate, turning the graph network from a bandwidth coordinator into a general channel-adaptive scheduler.","The OPV2V benchmark uses simulated CARLA scenes; testing on real-world V2V traces would clarify whether the reported masks and keep ratios survive real sensor noise and pose error."],"forward_implications":["If the reported figures hold, collaborative perception systems can operate at V2V bandwidths below 2 MB/frame on OPV2V-like scenes without dropping below state-of-the-art detection accuracy.","The two-stage design separates content selection from per-agent budget assignment, implying the same framework can be dropped onto any BEV backbone, since the reduction operates on confidence maps rather than the raw sensor stream.","The transmission-rate feedback term in the keep-ratio equation gives a concrete mechanism for congestion-aware operation: as the instantaneous rate rises, every vehicle automatically prunes more aggressively.","Because the selective-transmission module is trained with random top-$K$ sparsity, the network is exposed to a range of sparsity patterns, which should make it robust to unseen communication budgets at deployment."],"supporting_citations":[{"why":"Supplies the OPV2V benchmark, its fusion pipeline, and the Culver City test split that all comparisons are evaluated on.","marker":"[5]"},{"why":"Defines the spatial-confidence masking baseline (Where2Comm) that EffiComm extends and the main efficiency point of comparison.","marker":"[14]"},{"why":"Provides the PointPillars backbone used for all methods so that performance gaps come from the communication scheme rather than the perception backbone.","marker":"[20]"},{"why":"Introduces the sparsely-gated mixture-of-experts layer that motivates the MoE fusion head's expert-routing design.","marker":"[18]"},{"why":"Shows scaling vision models with sparse mixture of experts, informing the choice of expert specialization in the fusion module.","marker":"[19]"},{"why":"Provides the rate-distortion theory used to justify reducing keep ratios when network congestion increases.","marker":"[22]"}],"fun_headline_variants":["EffiComm cuts V2V data by more than 60% while keeping top 3D accuracy","EffiComm sends under 40% of prior data, keeping top 3D accuracy","Under 40% bytes, EffiComm still achieves SOTA 3D detection","EffiComm's two-stage pruning sends only useful BEV cells at 1.5 MB/frame","Learning which BEV cells to air hits SOTA with far fewer bytes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported communication savings are computed from the $\\log_2$ of the number of non-zero elements in the transmitted BEV features, and the paper does not specify how sparse indices, quantization, and per-vehicle mask overhead translate into actual bytes on the air.","fun_headline_variants_meta":{"raw":{"variants":["EffiComm cuts V2V data by more than 60% while keeping top 3D accuracy","EffiComm sends under 40% of prior data, keeping top 3D accuracy","Under 40% bytes, EffiComm still achieves SOTA 3D detection","EffiComm's two-stage pruning sends only useful BEV cells at 1.5 MB/frame","Learning which BEV cells to air hits SOTA with far fewer bytes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002027,"raw_usage":{"total_tokens":7911,"prompt_tokens":968,"completion_tokens":6943,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":6826}},"tokens_in":584,"tokens_out":6943,"duration_ms":44277,"temperature":1.0,"reasoning_tokens":6826,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:54:21.685806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument a full sparse-encoding version of EffiComm—index lists for kept cells, quantized feature values, per-vehicle keep ratios, and mask signaling—and transmit over a V2V channel emulator on the OPV2V test set. If the measured mean bytes per frame is not below 40% of a similarly instrumented Where2Comm run while keeping AP@0.7 within the reported margin, the paper's central efficiency claim is wrong.","supporting_citations":[{"cited_title":"Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication,","cited_arxiv_id":null,"evidence_quote":"Supplies the OPV2V benchmark, its fusion pipeline, and the Culver City test split that all comparisons are evaluated on."},{"cited_title":"Where2comm: Communication-efficient collaborative perception via spatial confi- dence maps,","cited_arxiv_id":null,"evidence_quote":"Defines the spatial-confidence masking baseline (Where2Comm) that EffiComm extends and the main efficiency point of comparison."},{"cited_title":"Outrageously large neural networks: The sparsely- gated mixture-of-experts layer,","cited_arxiv_id":null,"evidence_quote":"Introduces the sparsely-gated mixture-of-experts layer that motivates the MoE fusion head's expert-routing design."},{"cited_title":"Scaling vision with sparse mixture of experts,","cited_arxiv_id":null,"evidence_quote":"Shows scaling vision models with sparse mixture of experts, informing the choice of expert specialization in the fusion module."},{"cited_title":"Rate distortion theory: A mathematical basis for data compression,","cited_arxiv_id":null,"evidence_quote":"Provides the rate-distortion theory used to justify reducing keep ratios when network congestion increases."}],"review_version":2}