{"id":"b0a17bed-5196-48ad-bf5d-466bd5177c16","arxiv_id":"2508.19638","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CoPLOT replaces BEV features with semantically ordered, frequency-enhanced point-level tokens for collaborative perception, improving 3D detection while cutting overhead.","lead":"This paper introduces CoPLOT, a collaborative perception framework that shares point-level 3D tokens between vehicles instead of 2D bird's-eye-view maps. It reports higher detection accuracy with lower communication and computation overhead on three datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Top-k token selection is the linchpin of CoPLOT's efficiency claims, but the paper never measures whether the transmitted tokens actually contain objects at inference; the k=1300 budget is derived from test-set averages and may drop objects in dense scenes.","rationale":"The reader's weakest assumption identifies semantic-importance-score generalization as the key risk. I agree that this is the central load-bearing component, but I sharpen it into a concrete, testable gap: the paper provides no object-level recall measurement for the selected top-k tokens, and the value of k itself is justified using test-set statistics in Supp. E.2. If the importance scores generalize poorly to dense or unusual scenes, or if k has been tuned on the test split, the reported accuracy and overhead advantages would not hold at deployment. This does not overturn the reader's CONDITIONAL verdict; it reinforces the condition that must be verified. The proposed check—measuring whether transmitted tokens actually cover ground-truth objects, stratified by scene density, and re-deriving the k-choice from validation—would settle the concern directly.","tokens_in":25247,"tokens_out":9149,"duration_ms":108026,"concrete_test":"On OPV2V validation, freeze the trained CoPLOT encoder and, for each frame with k=1300, compute the recall of ground-truth objects among transmitted tokens (a box counts as recalled if at least one selected token lies inside it). Stratify by scene vehicle count (e.g., ≤10, 11–20, 21–33). If object recall drops materially with density or is below ~0.95 overall, the fixed top-k cannot support the headline trade-off. Additionally, re-run the Fig. 6 curve with k chosen on validation only, not test, and compare AP@0.7.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy/efficiency trade-off rests on the semantic-importance scores in §3.2: after focal-loss supervision with ground-truth boxes (Eq. 3), the encoder ranks tokens and transmits only the top-k. The paper never directly measures whether those selected top-k tokens actually cover the objects at inference. The fixed k=1300 used for OPV2V is justified in Supp. E.2 by averaging over the test set (~15 vehicles × 86.7 tokens/vehicle), but the same section reports up to 33 vehicles per scene; an average-based k will shed foreground tokens in dense scenes. If the importance head drops an object under occlusions or unusual geometry, the 'compact yet comprehensive' token sequence is not comprehensive, and the claimed AP gains plus ~90% communication reduction are overstated for those scenes. The post-hoc test-set peak also raises the possibility that k was chosen with test labels rather than validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoPLOT, a collaborative perception framework that replaces 2D BEV intermediate features with point-level tokens. The method introduces a point tokenizer, a semantic-aware token reordering module, a frequency-enhanced state space model, and a neighbor-to-ego alignment module for multi-agent fusion. Experiments on OPV2V, V2V4Real, and DAIR-V2X report state-of-the-art AP@0.5/0.7 for collaborative 3D detection, together with reduced computational and communication overhead. The supplementary material provides component ablations, reordering comparisons, and efficiency analyses.","tokens_in":25531,"tokens_out":5239,"duration_ms":58070,"significance":"If the claims hold, CoPLOT offers a genuine alternative to BEV-based intermediate fusion, preserving 3D structure while achieving strong efficiency. The paper's strengths are its three-dataset evaluation, a broad set of baselines, per-module ablations, and a candid limitations section. However, the central efficiency/accuracy trade-off rests on the top-k token selection mechanism, and this mechanism is not directly validated for object coverage at inference; additionally, the token budget k is justified using post-hoc test-set statistics. These issues are load-bearing but addressable, so the contribution is potentially significant pending verification.","major_comments":[{"comment":"The top-k token selection is the linchpin of CoPLOT's efficiency claims, but the paper never measures whether the transmitted tokens actually contain foreground objects at inference. The semantic-importance head is supervised with ground-truth box labels (Eq. 3), so it is a learned predictor; no experiment reports object recall or foreground-token coverage under the fixed k budget. The chosen k=1300 for OPV2V is justified in Supp. E.2 using test-set average statistics (15 vehicles × 86.7 tokens per vehicle), while the same section reports up to 33 vehicles per scene. An average-based k will therefore shed foreground tokens in dense scenes, and the claimed accuracy/efficiency trade-off may be optimistic for those frames. Please add validation-set experiments measuring the fraction of GT objects with at least one selected token, report AP under scene-density buckets, and select/report k ba","section":"§3.2, Eq. (3); Supp. E.2"},{"comment":"The explanation for the performance peak is post-hoc: the projected-area calculation uses a hand-picked 2/3 visibility factor and fixed vehicle dimensions, with no sensitivity analysis or predictive validation. The authors themselves note the calculation is based on average test statistics and that the optimal k fluctuates. This cannot serve as a principled justification for the fixed threshold. At minimum, reframe the calculation as a heuristic intuition and provide a validation-based sensitivity analysis over the area factor, vehicle size, and density assumptions.","section":"Supp. E.2"},{"comment":"The headline efficiency reductions ('approximately 80% and 90%') are stated without uniform reference baselines and without error bars. Table 4 shows CoPLOT at 176.2 GFLOPs versus CollaMamba at 198.3 GFLOPs, a reduction of only ~11%, while the 80% figure is relative to the heaviest baselines such as V2VNet. Communication reductions are only presented graphically (Fig. 4b) without exact byte counts, and no variance across runs or seeds is reported. Because computation and communication savings are central claims, please report mean±std over multiple runs, specify the reference baseline for each percentage, and tabulate actual message sizes (tokens + coordinates + pose) for all compared methods.","section":"§4.2, Table 4, Fig. 4b; Abstract"}],"minor_comments":[{"comment":"Typos: 'Samentic-aware' should be 'Semantic-aware'.","section":"§3.2 title"},{"comment":"The x-axis labels read 'Num of neb tokens' while the captions say 'point-level tokens'; please unify the terminology.","section":"Supp. E.2 / Fig. 5-6"},{"comment":"The loss Ls for semantic importance is defined but not shown in the total training loss. Specify how Ls, Loff, and detection losses are combined and weighted.","section":"§3.2, Eq. (3)"},{"comment":"CollaMamba and CoMamba are cited as arXiv preprints; if later peer-reviewed versions exist, cite them for reproducibility.","section":"References"},{"comment":"Grid-interval sensitivity is evaluated only on OPV2V. Reporting the same analysis on at least one real-world dataset (e.g., V2V4Real) would strengthen the choice of 0.4 m, since real-world point density differs considerably from simulated data.","section":"Supp. D.2, Table 6"},{"comment":"The claim that downsampling rate 4 is the best balance between performance and cost is weakly supported: for V2V4Real AP@0.7, 4× gives 0.447 vs 0.446 for 2×, while 1× gives 0.430. The performance differences are small; please add statistical significance or additional metrics to justify the chosen rate.","section":"Supp. D.4, Table 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a top-tier vision/robotics venue. The main technical risk is the test-set-informed selection of the top-k threshold and the lack of direct validation that selected tokens cover objects at inference. These issues are fixable with additional experiments and would not require a change in the method's overall architecture. I would support acceptance after the authors address the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CoPLOT makes a genuine break with the BEV consensus: instead of compressing multi-agent point clouds to a 2D grid, it exchanges point-level tokens and processes them with a Mamba-style sequence model. That is the paper's real contribution, and it is well executed. The architecture is coherent, the ablations support each component, and the supplementary is unusually thorough, including efficiency tables and a candid limitations section that admits the static top-k threshold and independent token selection are weaknesses.\n\nThe soft spots are real but not fatal. The biggest is the top-k token selection. The semantic-importance scores are trained with ground-truth boxes in the loop, and the fixed k=1300 used for OPV2V is derived, as the paper admits, from test-set averages (Supp E.2). The stress-test concern lands: with up to 33 vehicles per scene, an average-derived k can drop foreground tokens, and the paper never directly measures whether transmitted tokens cover objects at inference. The authors acknowledge this in the limitations and even suggest a dynamic threshold as future work, but the efficiency claims (80% compute, 90% communication reduction) rest on this mechanism for the test set, and the lack of a validation-based k selection is a methodological gap.\n\nTwo smaller points: no error bars or significance tests make the 4.2%, 5.3%, 9.8% margins look firmer than they are; and the code isn't out yet, so the efficiency numbers (Table 4) can't be independently checked. The frequency-enhancement motivation leans on biological vision analogies, but the math in the supplement (Eq. 18-19) is concrete and doesn't depend on the analogy.\n\nNone of this breaks the central claim. The representation change away from BEV is a real idea, and the experiments are broad enough that the paper deserves a serious referee. The main thing a referee should demand: a separate validation-based selection of k, a robustness breakdown by scene density, and ideally error bars. If those hold, this is a solid advance for the V2X collaborative-perception subfield.\n\nI'd bring it to reading group and cite it if I were working in this area. Send it to peer review.","headline":"CoPLOT makes a real break from BEV by exchanging point-level tokens, but the fixed top-k is tuned on the test set and the efficiency numbers hinge on it.","tokens_in":25977,"tokens_out":2999,"would_cite":true,"duration_ms":30876,"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":"CoPLOT swaps BEV feature maps for compact point-level tokens in collaborative 3D detection, beating the next-best method by 4.2–9.8% AP@0.7 on three benchmarks while using ~80% less computation and ~90% less bandwidth.","keywords":["collaborative perception","3D object detection","point-level tokens","state space models","bird's-eye-view representation","LiDAR point clouds","vehicle-to-everything (V2X)","token reordering"],"falsifier":"A concrete test: keep CoPLOT's importance scorer fixed and raise object density well beyond the training distribution (e.g., the densest OPV2V scenes with 33 vehicles) without raising k; if per-object foreground recall collapses and AP@0.7 follows, the claim that the compact top-k sequence preserves objects fails. A second test targets generalization across gaps: run the scorer trained on simulated OPV2V directly on V2V4Real frames; a large AP drop would mean the scores overfit the simulated geometry rather than capturing objectness itself.","tokens_in":25206,"feed_emoji":"🚗","tokens_out":11507,"duration_ms":106312,"temperature":0.7,"pith_summary":"CoPLOT claims that the standard currency of collaborative perception — 2D bird's-eye-view feature maps exchanged between vehicles — throws away exactly the 3D structural detail needed to find and localize objects, and that point-level tokens are a better currency. The paper builds a point-native pipeline that turns LiDAR into a compact 1D token sequence: tokens are reordered by semantic group, modeled by a frequency-enhanced state space model, and the most object-relevant top-k are transmitted, then spatially realigned at the receiving agent. On the OPV2V, V2V4Real, and DAIR-V2X benchmarks it reports out-detecting every compared method, with the largest margins on real-world data (up to 9.8% AP@0.7), while consuming roughly 80% less computation and 90% less communication. If the claim holds, it redirects a key design choice: instead of compressing and sending dense 2D grids, agents should send a few thousand semantically ordered 3D tokens, which also makes collaborative perception compatible with sequence models and tight bandwidth budgets.","feed_headline":"Point tokens beat BEV maps for car-to-car 3D detection","feed_subtitle":"CoPLOT slashes compute ~80% and bandwidth ~90% while out-detecting prior collaborative perception models.","key_machinery":"The central mechanism is the point-level token sequence together with three modules that make a 1D token stream viable for outdoor LiDAR. (1) Semantic-aware token reordering (STR) uses scene-level and token-level prompts to assign each token a semantic group index, so semantically related tokens become adjacent in the serialized sequence. (2) The frequency-enhanced state space model (FSSM) is a Mamba-style selective SSM whose output matrix C — formally the counterpart of the query matrix in linear attention — is augmented with a compact frequency descriptor (DC energy, low- and high-frequency energy, and their ratio, from a windowed 2D DFT), letting the model separate object contours from br","core_discovery":"On the paper's own terms, the discovery is that point-level tokens — fixed-grid aggregations of raw LiDAR points that keep spatial distribution, intensity, density, and sensor distance — can carry enough structural and semantic information to serve as the shared message in multi-agent 3D detection, and that a pipeline of three components makes this work. Semantic-aware token reordering serializes the unordered 3D tokens into a 1D sequence so tokens that refer to the same or neighboring objects sit together, which a Mamba-style state space model can then process with linear complexity. A frequency-enhanced state space model adds a per-token spectral descriptor, derived from a local 2D DFT of","pith_inferences":["A scene-adaptive top-k is the natural next step, and the paper's own analysis points to it: the ~1300-token optimum on OPV2V matches roughly 15 vehicles × ~87 tokens per vehicle, so a module that predicts how many objects a scene holds could beat any fixed threshold; the paper lists the static threshold as a limitation but does not build this.","Because the importance scores are trained from ground-truth box containment, CoPLOT inherits the closed-world object vocabulary of its training set; an untested extension is to supervise importance with self-supervised foreground cues (free-space, occupancy, temporal persistence) so the compact token stream survives novel object classes.","The 1D, semantically ordered token stream is shaped by design for sequence backbones such as LLMs; the paper states this compatibility as motivation but does not test it, so a concrete extension is feeding CoPLOT's selected tokens into a language-model-style head for open-vocabulary or language-grounded queries.","Since the tokens retain dense 3D structure to the final stage, the same encoder and fusion pipeline should transfer to 3D occupancy prediction or semantic segmentation with only a head swap — a testable claim the paper leaves open."],"forward_implications":["If CoPLOT is right, the design space of collaborative perception shifts from compressing 2D BEV grids to serializing, ranking, and aligning sparser 3D token streams — the reported ~90% communication reduction comes from transmitting only object-relevant tokens.","The reported gains are largest on the two real-world datasets (V2V4Real and DAIR-V2X, +5.3% and +9.8% AP@0.7), where localization noise and height compression are most severe — consistent with the claim that point tokens preserve exactly the cues BEV pipelines damage.","Because the token stream is a genuine 1D sequence, CoPLOT makes state space models a viable backbone for outdoor LiDAR collaboration: the frequency injection recovers foreground/background discrimination that plain Mamba lacks, at linear complexity.","The relationship between the number of selected tokens and accuracy is a tunable resource curve — computation and communication scale linearly with the selected token count — so a deployed system can adaptively trade perception quality for latency, energy, and bandwidth.","The noise-sweep experiments show CoPLOT retains accuracy where BEV-based baselines degrade, indicating that explicit learned alignment of token positions is a workable substitute for perfect pose knowledge in real deployments."],"supporting_citations":[{"why":"OPV2V, the simulated V2V benchmark that supplies most of the training data and the primary comparison protocol.","marker":"(Xu et al. 2022c)"},{"why":"V2V4Real, the real-world dataset whose localization noise and domain shift drive the alignment and robustness experiments.","marker":"(Xu et al. 2023)"},{"why":"DAIR-V2X, the real-world vehicle-infrastructure benchmark where CoPLOT reports its largest margin.","marker":"(Yu et al. 2022)"},{"why":"Where2comm, the spatial-confidence BEV selection baseline the paper extends and must beat at matched budgets.","marker":"(Hu et al. 2022)"},{"why":"CollaMamba, the closest existing SSM-based collaborative perception method and a key baseline that still operates on 2D features.","marker":"(Li et al. 2024b)"},{"why":"Supplies the linear-attention/state-space correspondence used to justify injecting frequency features into the SSM's output matrix C.","marker":"(Han et al. 2024a)"},{"why":"The frequency-domain camouflaged-object detection work that motivates frequency-selective filtering for foreground/background separation.","marker":"(Zhong et al. 2022)"},{"why":"V2X-ViT, the transformer-based fusion baseline that defines the computational-overhead comparison point.","marker":"(Xu et al. 2022b)"}],"fun_headline_variants":["Point tokens beat BEV in collaborative 3D detection","CoPLOT delivers point tokens for faster, sharper multi-agent perception","Point-level tokens outperform BEV maps in vehicle-to-vehicle perception","Semantic reordering unlocks point tokens for collaborative perception","Cut bandwidth and compute with point-token collaborative perception"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Everything hinges on the learned \"which tokens matter\" scores: they are trained only from ground-truth object boxes, the agents transmit only the top-k tokens, and if those scores misrank objects in unfamiliar scenes, both the accuracy gain and the ~90% bandwidth saving collapse.","fun_headline_variants_meta":{"raw":{"variants":["Point tokens beat BEV in collaborative 3D detection","CoPLOT delivers point tokens for faster, sharper multi-agent perception","Point-level tokens outperform BEV maps in vehicle-to-vehicle perception","Semantic reordering unlocks point tokens for collaborative perception","Cut bandwidth and compute with point-token collaborative perception"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000414,"raw_usage":{"total_tokens":1994,"prompt_tokens":781,"completion_tokens":1213,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":1130}},"tokens_in":525,"tokens_out":1213,"duration_ms":12993,"temperature":1.0,"reasoning_tokens":1130,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:35:21.423262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: keep CoPLOT's importance scorer fixed and raise object density well beyond the training distribution (e.g., the densest OPV2V scenes with 33 vehicles) without raising k; if per-object foreground recall collapses and AP@0.7 follows, the claim that the compact top-k sequence preserves objects fails. A second test targets generalization across gaps: run the scorer trained on simulated OPV2V directly on V2V4Real frames; a large AP drop would mean the scores overfit the simulated geometry rather than capturing objectness itself.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The frequency-domain camouflaged-object detection work that motivates frequency-selective filtering for foreground/background separation."}],"review_version":1}