{"id":"44f2ec8b-dc67-4518-969a-2c5179c00b16","arxiv_id":"2505.00740","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Fast2comm improves collaborative perception accuracy and bandwidth efficiency by supervising the confidence map with ground-truth labels and using bounding-box priors during training, yet the claimed test-time benefits rely on a train-test mismatch.","lead":"Fast2comm is a collaborative perception method that uses ground-truth bounding boxes during training to select which features vehicles share, and a confidence map to choose features at test time. It reports better accuracy and bandwidth trade-offs than Where2comm on three datasets, but the test-time selection mechanism is not fully specified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time inference removes the GT-Bbox branch while fusion was trained on its output, so the reported gains may not transfer to the deployed configuration.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the training-time GT-Bbox selection is not used at test, and the fusion module receives a different input distribution. My stress-test confirms this is the single most consequential issue in the paper. It directly threatens the central claim because Table I and Figure 6 report the performance of the test-time configuration, yet the architecture description, ablation, and attribution of robustness all rely on the training-time G branch. The paper provides no analysis of how the confidence features M alone reproduce the selection behavior learned with G, and no details of how the self-attention handles the missing branch. The concern is not resolved by the available code link because the paper does not state what the code does at inference; a direct code trace and the three-arm experiment would either confirm the discrepancy or reveal that an undocumented zero-padding or masking mechanism makes the train/test gap harmless. Since the reader's conditional verdict already captures this uncertainty, no change to the verdict is needed; the condition should be that the authors supply the missing implementation specification and validation.","tokens_in":11133,"tokens_out":3555,"duration_ms":43596,"concrete_test":"Run the released inference code and trace the exact tensor passed to the Self-Attention module when only M is shared: determine whether G is omitted, zero-padded, or masked. Then evaluate three variants on the OPV2V validation set with identical checkpoints where possible: (a) train and test with both M and G (using oracle GT boxes at test), (b) train and test with M only, and (c) the paper's configuration: train with M+G, test with M only. If (c) is close to (b), the reported advantage does not come from the claimed test-time mechanism; if (c) is close to (a), the missing specification is a documentation issue. Reporting AP@0.7 and communication bytes for all three will settle whether the train/test decoupling is benign.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Fast2comm delivers a superior perception-communication trade-off and localization robustness. But Section III.E states: 'Fast2comm shares both prior features G and confidence features M during training, but only shares confidence features M during testing.' The fusion module defined in Eq. 5 is trained on the concatenation of M and G via self-attention; at test time, G is absent, and the paper does not specify how the self-attention input is re-formed (zero padding, masking, or length change). Since the self-attention and decoder never observed a sequence containing only M during training, their behavior on such inputs is untested. More importantly, the claimed bandwidth and robustness benefits of the GT Bbox-Based Feature Selection module in Sections I and IV are directly tied to selecting and sharing G, but G is not shared at inference. Thus the test-time system is effectively a different architecture from the one whose ablations and gains are reported. The ablation in Table II makes this especially acute: GT-FS alone degrades AP@0.7 on OPV2V from 78.44 to 71.82, and only the CFG+GT-FS combination exceeds baseline; this suggests the reported improvements depend on a training-only interaction that is not demonstrated to survive the removal of G at test time.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Fast2comm, a collaborative perception framework for multi-agent 3D object detection that combines prior-supervised confidence feature generation, GT bounding box-based feature selection, and a decoupled training/testing fusion strategy. The authors claim that Fast2comm achieves a better perception-communication trade-off and improved robustness to localization errors compared with prior methods, with reported gains over Where2comm of 1.0%/1.2% (OPV2V), 2.7%/2.9% (V2XSet), and 1.5%/0.9% (DAIR-V2X) in AP@0.5/0.7. The manuscript includes experiments on three datasets, ablation studies, communication-volume comparisons, and qualitative visualizations.","tokens_in":11354,"tokens_out":2907,"duration_ms":33342,"significance":"If the reported results hold at inference time, the idea of using ground-truth bounding boxes as a training-time structural prior to shape a sparse confidence-based communication channel is a useful contribution to collaborative perception, and the public code release is a positive step. The empirical study is broad, covering three datasets and including robustness and bandwidth analyses. However, the central claim is undermined by a train/test discrepancy: the fusion module is trained on concatenated confidence and GT-bbox features but at test time only confidence features are shared, and the paper does not specify how the trained fusion module handles the missing input. Because the reported gains are obtained in a configuration that is not the deployed one, the experimental validation as written does not establish the paper's main claims.","major_comments":[{"comment":"The paper states that during training both prior features G and confidence features M are shared, but during testing only M is shared. The fusion module in Eq. (5) applies self-attention to the flattened concatenation of M and G, so the trained model has never observed inputs containing only M during training. The manuscript does not specify how the self-attention input is re-formed at test time (zero-padding, masking, or a modified sequence length), nor does it report results for the test-time configuration. Since all quantitative comparisons in Tables I and Figures 5-6 are presumably obtained with a model that at test time receives only M, the reported accuracy and bandwidth figures cannot be taken at face value without a description and validation of this adaptation.","section":"Section III.D, III.E, Eq. (5)"},{"comment":"The robustness improvement in Figure 6 is attributed to both the Confidence Feature Generation module and the GT Bbox-Based Feature Selection module, but GT-FS is a training-only mechanism: at test time the GT bbox-based branch is not shared. The ablation in Table II shows that GT-FS alone degrades OPV2V AP@0.7 from 78.44 to 71.82, and only the combination CFG+GT-FS exceeds the baseline. This combination is trained with both M and G, yet at test time G is absent; the manuscript does not demonstrate that the trained model retains this improvement when G is removed. The robustness claim therefore conflates training-time supervision with test-time feature selection, and the experiment needs to be rerun or explicitly reported in the actual test-time configuration.","section":"Section IV.C.3, IV.E, Table II"},{"comment":"The ablation study is underspecified regarding the training and testing protocol. It is unclear whether each ablation row uses the same training-time inputs (M and G) and only changes test-time inputs, or whether the presence of GT-FS changes the training procedure itself. In particular, the row with CFG only reports lower AP than the baseline, but if the baseline already used only M at test time, the comparison is not controlled for the fusion input. The authors should clarify how the train/test decoupling applies in each ablation row and provide results for the exact deployed setting (training with M and G, testing with M only) with appropriate architectural adaptation.","section":"Section IV.E, Table II"}],"minor_comments":[{"comment":"Typos and inconsistent naming appear in several places: 'F ast2comm' in the abstract, 'Faste2comm' in Section III.B.2, 'reprensents' in Eq. (2), 'Boudning' in the Section III.C heading, 'Robutness' in the Section IV.C.3 heading, 'hte' in Section III.G, and 'Where2Comm' with inconsistent capitalization. A careful proofread is needed.","section":"Throughout"},{"comment":"The confidence map threshold t in Eq. (1) is never defined or discussed in the experiments. The value of t directly controls the sparsity of M and hence the communication volume, so the authors should report the threshold used in all experiments and whether it is tuned per dataset.","section":"Section III.B.2"},{"comment":"In Table I, the 'No Fusion' and 'Early Fusion' rows report identical numbers on all three datasets; this is likely a typo or a copy-paste error and should be corrected or explained.","section":"Section IV.A, Table I"},{"comment":"Figures 5 and 6 are referenced only with minimal captions and are not discussed in the body text with quantitative details; the authors should add explicit descriptions of what is plotted and which curves correspond to which setting.","section":"Section IV.D"},{"comment":"The loss formulation uses smooth L1 for regression and focal loss for both classification and prior knowledge supervision, but the weights alpha, beta, and gamma are not specified in the text. Please report the chosen values and whether they were tuned.","section":"Section III.G"}],"recommendation":"major_revision","confidential_remarks":"The central concern is that the paper evaluates a configuration that differs from the deployed test-time system, and this gap is not acknowledged or analyzed. The novelty over Where2comm is incremental (adding a GT-supervised confidence map and a training-only bbox selection), but the prior-supervised confidence generation could be a viable direction if the test-time behavior is carefully characterized. I would like the editor to weigh whether the authors can provide the missing test-time validation within a revision; if they cannot, the reported gains should be treated as unsubstantiated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the Fast2comm paper. The short version: the core idea is neat, but the paper doesn't currently demonstrate that the reported gains survive the test-time setup.\n\nWhat's actually new: they supervise the confidence map with ground-truth labels to make it sharper, and they use GT bounding boxes to select features during training only. At test, they only share the confidence-gated features M, not the GT-selected features G. That decoupling is the paper's main trick, and it's plausible—train with richer supervision, test with a cheaper message. The experiments span three datasets, the code is public, and the numbers beat Where2comm consistently.\n\nThe problem is the test-time behavior is under-specified. The fusion module (Eq. 5) is trained on concatenated M and G, but at test G is gone. The paper never says whether G is zero-padded, the sequence length changes, or the self-attention somehow copes with missing inputs. Since self-attention is length-agnostic but feature-dimension-sensitive, this matters a lot. If G is concatenated along channels, the model sees a different input dimension; if along spatial tokens, it sees half the context. Either way, the test-time input distribution differs from training. The authors might have handled this with a masking trick, but it's not in the paper.\n\nThe robustness claim is also partly tied to GT-FS, which isn't used at test. Section IV.C attributes the localization-robustness gains to both CFG and GT-FS, but GT-FS doesn't exist at inference. Those gains must come from the improved confidence map alone, and the ablation doesn't isolate that cleanly. In fact, Table II shows that adding GT-FS alone hurts a lot on OPV2V (78.44 to 71.82), which suggests the interaction is fragile.\n\nThe ablation is the most unusual part. CFG alone barely changes, GT-FS alone hurts, both together help. That's not intrinsically wrong—synergies happen—but the explanation in the paper is hand-wavy, and it doesn't address what happens at test when G is absent.\n\nAll that said, I don't think the paper is a wash. The prior-supervised confidence map is a sensible idea, and the reported improvements might be real, but the current write-up doesn't prove the test-time transfer. A good revision would specify the inference-time fusion, add an ablation where G is dropped at test, and retract the robustness attribution if it doesn't hold.\n\nI'd send this to peer review rather than desk reject—it's a credible incremental contribution with code and multi-dataset evaluation—but the reviewers should focus on closing the train/test gap.","headline":"GT-supervised feature selection is a reasonable idea, but the train/test decoupling is under-specified and the robustness claim needs better support.","tokens_in":11879,"tokens_out":3441,"would_cite":false,"duration_ms":35519,"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":"Fast2comm claims that collaborative perception improves when ground-truth bounding boxes supervise the confidence features that agents share, and that sharing fewer, target-gated features beats sharing top-k confidence features.","keywords":["collaborative perception","multi-agent perception","V2X autonomous driving","BEV feature selection","confidence map","bandwidth efficiency","localization robustness","3D object detection"],"falsifier":"Train the same Fast2comm with the GT-box-gated features also shared at test time and measure AP; compare it with the paper's reported test-time setting that shares only confidence-gated features. If the accuracy drop from removing the GT-gated features is large, the reported gains depend on a training-only signal that real inference does not have.","tokens_in":10885,"feed_emoji":"🚗","tokens_out":6570,"duration_ms":64227,"temperature":0.7,"pith_summary":"The paper proposes Fast2comm, a collaborative perception framework in which vehicles share a sparse set of Bird's Eye View features chosen by prior knowledge. It claims that supervising a confidence map with ground-truth labels, and using ground-truth bounding boxes to select which feature regions are shared, gives a better accuracy-bandwidth trade-off than the top-k confidence selection used by Where2comm. Across OPV2V, V2XSet, and DAIR-V2X, Fast2comm reports consistent gains over Where2comm, with the largest being 2.7 and 2.9 AP points on V2XSet, and it keeps an advantage when localization noise is added. If the claim holds, connected vehicles could transmit less data while detecting occluded or distant objects more reliably.","feed_headline":"Fast2comm lifts collaborative detection up to 2.9 points on V2XSet","feed_subtitle":"Ground-truth boxes pick which regions cars share, keeping accuracy high while cutting transmitted bytes.","key_machinery":"The mechanism is a pair of feature-gating maps applied to the BEV feature map. The confidence map is produced by a detection head on an attention-fused feature representation, with an extra loss that supervises it against ground-truth labels so that it marks foreground. The second map is built by projecting each 7D ground-truth box into a 4D BEV box and binarizing it, so that it marks exactly the target regions. Both maps are applied by element-wise multiplication; agents transmit only non-zero entries, and the receiver concatenates the two gated tensors and fuses them with self-attention. At test time only the confidence-gated tensor is shared.","core_discovery":"The central claim is that ground truth can be made to do double duty: it trains a confidence head to separate foreground from background, and it defines a binary BEV mask that restricts shared features to target regions. The paper shows that these two mechanisms are coupled—each alone hurts performance relative to the baseline, while together they improve it—and argues this is because the confidence map tells the mask where objects are, while the mask keeps background noise out of the fusion. It also claims that the decoupled fusion strategy (sharing both feature types during training, only the confidence-gated ones at test) preserves accuracy while cutting communication cost.","pith_inferences":["Editorial inference: the decoupled training/testing design suggests a cheap extension—at test time, select a connected-region mask from the confidence map instead of per-pixel top-k, which may recover part of the GT-gated signal without needing labels.","Editorial inference: the paper's robustness gain under GPS noise is consistent with the selection mask being anchored to object-sized regions rather than isolated high-confidence pixels; if so, any collaborative perception method that shares spatially coherent masks should show similar noise tolerance.","Editorial inference: the reported degradation when only one module is used implies that prior-supervised confidence generation and GT-box selection are coupled in the loss landscape; a practitioner adopting one without the other should expect a performance drop."],"forward_implications":["Fast2comm reports gains over Where2comm of 1.0/1.2 AP points on OPV2V, 2.7/2.9 on V2XSet, and 1.5/0.9 on DAIR-V2X, with sparser messages.","Under simulated localization noise, the method's advantage over Where2comm grows (3.6 AP points at AP@0.5 with 0.5 m noise on OPV2V), indicating the prior-supervised confidence map helps absorb misalignment.","Because only non-zero selected features and their indices are transmitted, communication cost scales with sparsity; the decoupled training/testing design makes test-time messages smaller than training-time messages.","Ablations show the two proposed modules are synergistic: adding either alone lowers AP@0.7 on OPV2V and V2XSet, while adding both raises it above the baseline."],"supporting_citations":[{"why":"The main baseline; supplies the top-k confidence-map sharing scheme that Fast2comm replaces and the numbers it improves on.","marker":"[7]"},{"why":"PointPillars encoder used as the shared backbone for all feature extraction in the framework.","marker":"[17]"},{"why":"Supplies the scaled dot-product attention and self-attention used in feature fusion.","marker":"[19]"},{"why":"Provides the OPV2V dataset and the AttFuse baseline used in benchmark comparisons.","marker":"[13]"},{"why":"Provides the V2XSet dataset and the V2X-ViT model compared as prior state of the art.","marker":"[2]"},{"why":"Provides the DAIR-V2X real-world dataset and its evaluation protocol.","marker":"[14]"},{"why":"Scope, a recent state-of-the-art baseline whose OPV2V, V2XSet, and DAIR-V2X results are compared in Table I.","marker":"[24]"},{"why":"MRCNet, a motion-aware collaborative perception baseline that Fast2comm compares against on OPV2V and V2XSet.","marker":"[8]"}],"fun_headline_variants":["Fast2comm: GT-guided sharing trims bytes, gains 2.9 points","Confidence-gated features lift V2X perception by 2.9","Prior knowledge sharpens collaborative perception, cuts cost","Decoupled fusion: high accuracy, low bandwidth for cars"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the confidence map learned under ground-truth supervision selects the same target regions at test time that ground-truth boxes selected during training, and that the fusion module, trained on both feature types, still works when only the confidence-gated features arrive.","fun_headline_variants_meta":{"raw":{"variants":["Fast2comm: GT-guided sharing trims bytes, gains 2.9 points","Confidence-gated features lift V2X perception by 2.9","Prior knowledge sharpens collaborative perception, cuts cost","Decoupled fusion: high accuracy, low bandwidth for cars"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000279,"raw_usage":{"total_tokens":1616,"prompt_tokens":861,"completion_tokens":755,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":680}},"tokens_in":477,"tokens_out":755,"duration_ms":7701,"temperature":1.0,"reasoning_tokens":680,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:08:00.470375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same Fast2comm with the GT-box-gated features also shared at test time and measure AP; compare it with the paper's reported test-time setting that shares only confidence-gated features. If the accuracy drop from removing the GT-gated features is large, the reported gains depend on a training-only signal that real inference does not have.","supporting_citations":[{"cited_title":"Where2comm: communication-efficient collaborative perception via spatial confi- dence maps,","cited_arxiv_id":null,"evidence_quote":"The main baseline; supplies the top-k confidence-map sharing scheme that Fast2comm replaces and the numbers it improves on."},{"cited_title":"Pointpillars: Fast encoders for object detection from point clouds,","cited_arxiv_id":null,"evidence_quote":"PointPillars encoder used as the shared backbone for all feature extraction in the framework."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Supplies the scaled dot-product attention and self-attention used in feature fusion."},{"cited_title":"Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication,","cited_arxiv_id":null,"evidence_quote":"Provides the OPV2V dataset and the AttFuse baseline used in benchmark comparisons."},{"cited_title":"V2x-vit: Vehicle-to-everything cooperative perception with vision trans- former,","cited_arxiv_id":null,"evidence_quote":"Provides the V2XSet dataset and the V2X-ViT model compared as prior state of the art."},{"cited_title":"Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the DAIR-V2X real-world dataset and its evaluation protocol."},{"cited_title":"Spatio-temporal domain awareness for multi- agent collaborative perception,","cited_arxiv_id":null,"evidence_quote":"Scope, a recent state-of-the-art baseline whose OPV2V, V2XSet, and DAIR-V2X results are compared in Table I."},{"cited_title":"Multi-agent collaborative perception via motion-aware robust communication network,","cited_arxiv_id":null,"evidence_quote":"MRCNet, a motion-aware collaborative perception baseline that Fast2comm compares against on OPV2V and V2XSet."}],"review_version":1}