{"id":"dfdf161c-44f4-42fa-b1b8-1ab71d16b2bb","arxiv_id":"2502.08169","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CoDynTrust gates shared vehicle features by a learned trust score derived from aleatoric and epistemic uncertainty, improving 3D detection under time delays, though its evaluation leaks test-set statistics.","lead":"This paper proposes CoDynTrust, a LiDAR-based collaborative perception system that uses per-region uncertainty scores to decide which features from other vehicles to trust when data arrive at different times. It reports better 3D detection accuracy than prior asynchronous fusion methods on three datasets, but its evaluation uses test-set statistics for normalization, which may inflate the reported gains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set statistics in Eq. (4) calibrate the trust modulus on the evaluation data; the reported APs are not clean held-out numbers, so the SOTA claim is unsupported as written.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing flaw: Eq. (4) uses test-set statistics in the normalization, which invalidates the evaluation as reported. I examined other possible objections but none is as decisive. The absence of error bars and the overclaim about planning/control are real weaknesses, but they are secondary; they do not by themselves overturn the detection results. The architecture is plausible and the code release is a positive, but the empirical foundation of the paper is compromised by the explicit test-set calibration. A held-out recalibration is a straightforward check that would decide whether the method still works. Until that is done, the paper's central claim that CoDynTrust achieves state-of-the-art asynchronous collaborative perception is not established. This does not require judging the authors' intentions; it is a protocol-level problem with the reported numbers. My verdict thus stays aligned with the reader's REJECT, i.e., no change to the reader's verdict.","tokens_in":12576,"tokens_out":5627,"duration_ms":49742,"concrete_test":"Re-run the full benchmark (Table I and Table II) with µu, σu, µs, and σs estimated only on positive samples from the training split (or a dedicated calibration/validation split), leaving all other settings fixed. Then compare the resulting AP@0.5 and AP@0.7 values with the published numbers on DAIR-V2X, V2XSet, and OPV2V, and check whether CoDynTrust still ranks above CoBEVFlow and Where2comm+SyncNet at 300ms and 500ms delays. If the AP shifts by more than roughly 0.5 points or the ranking changes, the reported state-of-the-art result is an artifact of test-set normalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that CoDynTrust is state-of-the-art under temporal asynchrony rests on Table I, and Table I is generated with a normalization (Eq. 4) whose constants are explicitly computed on the test set: the paper states that µu, σu, µs, and σs are 'the mean and variance of raw classification uncertainty and confidence scores for positive samples in the test set.' These constants are used to rescale every ROI's classification uncertainty before the DFTM is produced, so the trust modulus is calibrated on the exact samples used for evaluation. This contaminates the empirical protocol: the model's feature weighting can adapt to the aggregate statistics of the test distribution, which is not available in a real deployment. The leakage is not a cosmetic detail; it affects every reported AP in Table I and Table II. The ablation in Table III does not isolate the effect, because the DFTM rows also use the leaky normalization. Without a re-run with calibration statistics derived from a held-out split, the SOTA claim is not supported. The architecture itself may be sound, and the provided code makes a re-check feasible, but the paper as written does not establish its main empirical conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoDynTrust, a collaborative perception framework for LiDAR-based 3D object detection under temporal asynchrony. The method quantifies aleatoric and epistemic uncertainty for each region of interest, rescales these uncertainties, and uses them with confidence scores to generate a dynamic feature trust modulus (DFTM) that weights features before asynchronous fusion. Motion compensation is done by linear extrapolation, and a multi-scale hybrid fusion module aggregates the weighted features. Experiments on DAIR-V2X, V2XSet, and OPV2V report state-of-the-art AP under expected delays from 0 ms to 500 ms, along with ablations and robustness tests under pose noise.","tokens_in":12774,"tokens_out":5937,"duration_ms":50871,"significance":"If the reported results are valid, the uncertainty-guided trust-modulus idea is a useful addition to asynchronous collaborative perception: it is architecturally simple, interpretable, and the released code would let others reproduce and extend the work. The paper also makes a reasonable move by propagating uncertainty to fusion rather than only to downstream planning. However, the empirical contribution is currently undermined by the normalization in Eq. (4), which explicitly uses test-set statistics computed from positive samples. This leaks label information into the evaluation and affects every reported AP in Tables I, II, and III. The significance of the paper therefore cannot be assessed from the numbers as presented; a clean re-run with validation-set normalization is required before the state-of-the-art claim is credible.","major_comments":[{"comment":"The rescaling operation in Eq. (4) uses µu, σu, µs, and σs, which the text defines as the mean and variance of raw classification uncertainty and confidence scores for positive samples in the test set. Because this rescaling is applied to every ROI before DFTM generation, the trust weights are calibrated using the evaluation split and, more seriously, using ground-truth positive/negative labels. This leaks test information into the model at inference and contaminates every AP in Tables I and II, as well as every row of the ablation in Table III. The state-of-the-art claim is therefore not supported by a clean held-out evaluation. The authors should recompute all normalization statistics from the training or validation split only, state the split explicitly, and rerun the experiments.","section":"§IV-B, Eq. (4)"},{"comment":"The ablation does not isolate whether the uncertainty input to DFTM is the source of the reported gains. Row 3 (Uncertainty + DFTM) adds both DFTM and the leaky rescaled uncertainty relative to Row 2, and Row 2 itself uses the same test-set normalization. There is no condition with confidence-only DFTM, and no condition that removes the test-set statistics while keeping the uncertainty model. Consequently, the improvement attributed to uncertainty could instead come from the confidence term or from adaptation to the test set. Additional controlled ablations are needed, at minimum confidence-only versus confidence-plus-uncertainty, both with and without the leaky normalization.","section":"Table III"},{"comment":"The claim of state-of-the-art performance is not uniform even in the presented table: on V2XSet at 0 ms expected delay, CoDynTrust achieves AP@0.5 = 0.8780, which is below Where2comm (0.8918) and V2X-ViT (0.8785). The abstract and conclusion state that CoDynTrust achieves state-of-the-art performance 'even with temporal asynchrony,' but this should be qualified to delay settings, or the zero-delay comparison should be discussed explicitly. This issue is secondary to the normalization problem but affects how the headline claim is stated.","section":"Table I"}],"minor_comments":[{"comment":"The table formatting is inconsistent: the 'Single' row appears to list three values only under the DAIR-V2X columns, with no corresponding entries for V2XSet or OPV2V. Please align the columns and clarify whether 'Single' is delay-independent or re-evaluated at each delay.","section":"Table I"},{"comment":"The left-hand side ucls and the raw input ucls_raw are notationally confusing; use a distinct symbol such as ucls_rescaled for the output of the rescaling operation.","section":"Eq. (4)"},{"comment":"The confidence score sc is used in Eq. (4) but is not defined in the text; please define it as the ROI classification confidence and specify how it is obtained.","section":"Eq. (4)"},{"comment":"There are several typos and formatting issues: 'Impelment Details' should be 'Implementation Details'; 'have be given' should be 'have been given'; the conclusion contains an incomplete sentence ('demonstrates strong robust'); and 'A VGOUT' should be 'AVGOUT'.","section":"II and V"},{"comment":"Reference [29] is attributed only to 'Kingma' but is the standard variational autoencoder paper by Kingma and Welling; the author list should be completed.","section":"References"},{"comment":"The notation in the problem formulation, such as Xti n n and Yti n n, is difficult to parse because subscripts and superscripts are not clearly distinguished; please clean up the typesetting.","section":"III"}],"recommendation":"major_revision","confidential_remarks":"The test-set normalization issue in Eq. (4) is confirmed by the manuscript text and is serious: it invalidates the reported numbers as clean held-out results. I am not recommending outright rejection because the architecture and code release are valuable and the problem is fixable by recomputing normalization statistics on a validation split and rerunning the experiments. If the authors cannot provide such a re-run, or if the improvements do not persist, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plain take: the architecture is worth a look, but the main empirical result is compromised. Eq. (4) normalizes classification uncertainty using the mean and variance of raw uncertainty and confidence for positive samples in the test set, and those constants flow directly into the DFTM that scales features before detection. That leaks test-set statistics into the evaluation, so the AP numbers in Tables I and II are not clean held-out results. The stress-test note is right, and the reader's skeptic is right to put this at the center.\n\nWhat's genuinely new: the paper combines per-ROI aleatoric/epistemic uncertainty with a learned trust modulus and a multi-scale hybrid fusion module, applying this to asynchronous collaborative perception. That is a real combination I haven't seen before. The design choices are mostly sensible—linear extrapolation for motion prediction avoids extra learned error, and the sparse feature exchange keeps communication cheap. The code is public, which makes a re-check feasible. On the plus side, the writing is clear about what the normalization does; this isn't hidden.\n\nSoft spots, in proportion: the test-set leakage is load-bearing, big. A re-run with statistics from a held-out validation split could fix it, but as written the SOTA claim is unsupported. The ablation in Table III doesn't isolate the uncertainty input to DFTM—comparing DFTM with and without UQ is not the same as ablating the uncertainty signal, since confidence alone may drive the gain. Also minor: the paper claims uncertainty can propagate to planning/control but never demonstrates that, and there are no error bars. The pose-noise experiment is a good addition but inherits the same leakage.\n\nNet: this is a plausible method with a fixable evaluation flaw, not a dead end. The paper deserves a serious referee because the idea is meaningful and the code is available, but it should be rejected in its current form and revised with clean calibration and sharper ablations. I'd bring it to reading group as a cautionary example of test-set normalization seeping into a learning pipeline.","headline":"A promising idea for uncertainty-gated asynchronous fusion is undermined by test-set normalization in Eq. (4); the SOTA claim is not supported as written, though the fix is straightforward.","tokens_in":569,"tokens_out":745,"would_cite":false,"duration_ms":24653,"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":"CoDynTrust — an asynchronous collaborative-perception framework — claims that per-region dynamic feature trust moduli, derived from aleatoric and epistemic uncertainty, suppress low-quality features and keep 3D detection accurate under…","keywords":["collaborative perception","temporal asynchrony","uncertainty quantification","dynamic feature trust modulus","3D object detection","V2X","LiDAR","feature fusion"],"falsifier":"Re-run the three dataset evaluations with the normalization statistics in Eq. (4) computed only from a training or validation split, then check whether CoDynTrust still beats CoBEVFlow at 300 ms and 500 ms expected delays by the margins in Table I; if the margins vanish, the state-of-the-art claim is an artifact of test-set normalization.","tokens_in":12366,"feed_emoji":"🚗","tokens_out":6293,"duration_ms":46486,"temperature":0.7,"pith_summary":"Temporal asynchrony—communication delay, clock skew, or different sampling rates—makes the features each agent receives misaligned in time, so fused perception degrades. CoDynTrust attacks this by giving every region of interest a dynamic feature trust modulus (DFTM): a number between 0 and 1 computed from the agent's own aleatoric and epistemic uncertainty, averaged over two historical frames, and decayed with delay. The modulus is scattered back onto the feature map before fusion, so low-confidence or noisy regions are suppressed rather than contaminating the shared representation. The paper reports that this keeps 3D detection accuracy close to the synchronized baseline at expected delays of 300–500 ms and outperforms prior latency-compensation methods on DAIR-V2X, V2XSet, and OPV2V. The broader point: uncertainty need not be an output of perception; it can be the mechanism that decides what gets fused.","feed_headline":"Per-ROI trust scores keep 3D detection accurate under V2X delays","feed_subtitle":"CoDynTrust suppresses stale, noisy features and beats prior methods at 300-500 ms latency on three benchmarks.","key_machinery":"The Dynamic Feature Trust Modulus (DFTM), a per-ROI scalar in [0,1] generated by a residual network from the averaged confidence and averaged aleatoric/epistemic uncertainties of the two most recent frames, multiplied by a delay decay factor exp(−0.02·Δt), is the central object. DFTM is scattered back onto the sparse BEV feature map and multiplies each ROI's features after the flow-based motion compensation, so features that are uncertain or stale are scaled down before multi-scale hybrid fusion. The complementary machinery is the uncertainty quantification itself: direct modeling for aleatoric uncertainty (Gaussian box center, von Mises yaw, Gaussian classification logit with the reparameterization trick) and Monte-Carlo dropout for epistemic uncertainty, unified by a rescaling step. The hybrid fusion module combines MAXOUT and AVGOUT branches with spatial and channel weighting to handle the sparse, scaled features DFTM produces.","core_discovery":"The central claim is that per-ROI uncertainty, properly rescaled and merged with confidence, is a sufficient signal to make asynchronous collaborative perception robust. To get that signal, CoDynTrust models aleatoric uncertainty by treating the bounding-box center as a Gaussian random variable, the yaw angle as von Mises distributed, and the classification logit as Gaussian, while epistemic uncertainty comes from T Monte-Carlo dropout inferences. After rescaling, averaged confidence and uncertainty feed a small network that outputs a trust modulus in [0,1]; a decay factor d_f = exp(−k·Δt) with k = 0.02 further lowers trust for older frames. The trust modulus is scattered back onto the sparse feature map, whose grids are shifted by a BEV flow map built from linear extrapolation of ROI positions over two frames. On DAIR-V2X, V2XSet, and OPV2V, the resulting pipeline retains more detection accuracy at 300–500 ms expected delays than V2VNet, V2X-ViT, Where2comm+SyncNet, and CoBEVFlow, and it stays ahead when Gaussian pose noise is added.","pith_inferences":["If the rescaling statistics in Eq. (4) must be computed on a validation set rather than the test set, the normalized uncertainty changes, and the reported AP margins are the thing to re-check; this is the paper's single most testable vulnerability.","A natural extension the paper does not explore is applying DFTM to tracking or prediction heads: a per-ROI trust signal could gate association or weigh trajectory forecasts, not just spatial features.","The linear-extrapolation assumption limits the approach to moderate, roughly constant-velocity motion; at longer delays or with maneuvering agents, an extension would need to compare the decay factor k=0.02 against learned delay-decay schedules.","Comparing CoDynTrust against a stronger motion model (e.g., the learned flow in CoBEVFlow) at delays beyond 500 ms would show whether the decay factor or the trust modulus is doing the heavy lifting."],"forward_implications":["At expected delays up to 500 ms, detection AP drops far less than with prior delay-compensation networks, so vehicles can fuse data from agents with different clock phases or congested channels without retraining.","Because the trust modulus is computed from confidence and uncertainty alone, the same mechanism applies to any ROI-based detector; the paper's ablation shows the gain comes specifically from combining uncertainty with DFTM rather than uncertainty alone.","The transmitted message is sparse features, ROIs, and uncertainty, which the bandwidth-vs-AP experiments show reaches strong detection performance at lower communication volume than Where2comm on V2XSet.","With pose noise up to 0.4 m / 0.4°, CoDynTrust remains ahead of CoBEVFlow, so the robustness is not bought at the cost of sensitivity to misalignment.","DFTM can be propagated to planning and control, since it is an interpretable per-object reliability score rather than a fixed network weight."],"supporting_citations":[{"why":"Provides the bidirectional ROI matching and BEV flow baseline that CoDynTrust replaces with linear extrapolation plus DFTM, and supplies the comparison that defines the delay-regime state of the art.","marker":"[15]"},{"why":"Supplies the classification deviation ratio used in Eq. (4) to rescale classification uncertainty onto a unified scale.","marker":"[24]"},{"why":"The Bayesian approximation through Monte-Carlo dropout that gives the epistemic uncertainty via T stochastic inferences.","marker":"[22]"},{"why":"Basis for direct modeling of aleatoric uncertainty by predicting variance along with the output.","marker":"[21]"},{"why":"The PointPillar encoder backbone that produces the BEV features in step 2a of the pipeline.","marker":"[1]"},{"why":"Real-world vehicle-infrastructure dataset used for the DAIR-V2X experiments and delay robustness evaluation.","marker":"[16]"},{"why":"V2XSet dataset and the V2X-ViT baseline that CoDynTrust is compared against under asynchronous settings.","marker":"[13]"},{"why":"OPV2V vehicle-to-vehicle dataset, specifically the Culver City Digital Twin test set used for generalization evaluation.","marker":"[17]"}],"fun_headline_variants":["CoDynTrust: per-ROI trust tames V2X time skew","Trust modulus from uncertainty counters V2X asynchrony","Per-ROI dynamic trust keeps lidar detection robust to delays","Uncertainty-driven trust fixes asynchronous collaborative perception"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The rescaling in Eq. (4) uses the mean and variance of raw uncertainty and confidence scores computed from positive samples in the test set; if those statistics instead have to come from a validation set (to avoid test-set leakage), the normalized uncertainties—and therefore the reported AP gains—would change.","fun_headline_variants_meta":{"raw":{"variants":["CoDynTrust: per-ROI trust tames V2X time skew","Trust modulus from uncertainty counters V2X asynchrony","Per-ROI dynamic trust keeps lidar detection robust to delays","Uncertainty-driven trust fixes asynchronous collaborative perception"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001028,"raw_usage":{"total_tokens":4378,"prompt_tokens":1034,"completion_tokens":3344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":3283}},"tokens_in":650,"tokens_out":3344,"duration_ms":25186,"temperature":1.0,"reasoning_tokens":3283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:10:17.650406+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the three dataset evaluations with the normalization statistics in Eq. (4) computed only from a training or validation split, then check whether CoDynTrust still beats CoBEVFlow at 300 ms and 500 ms expected delays by the margins in Table I; if the margins vanish, the state-of-the-art claim is an artifact of test-set normalization.","supporting_citations":[{"cited_title":"Asynchrony-robust collaborative perception via bird’s eye view flow,","cited_arxiv_id":null,"evidence_quote":"Provides the bidirectional ROI matching and BEV flow baseline that CoDynTrust replaces with linear extrapolation plus DFTM, and supplies the comparison that defines the delay-regime state of the art."},{"cited_title":"Uncertainty-encoded multi-modal fusion for robust object detection in autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Supplies the classification deviation ratio used in Eq. (4) to rescale classification uncertainty onto a unified scale."},{"cited_title":"Uncertainty in deep learning,","cited_arxiv_id":null,"evidence_quote":"Basis for direct modeling of aleatoric uncertainty by predicting variance along with the output."},{"cited_title":"Pointpillars: Fast encoders for object detection from point clouds,","cited_arxiv_id":null,"evidence_quote":"The PointPillar encoder backbone that produces the BEV features in step 2a of the pipeline."},{"cited_title":"Dair-v2x: A large-scale dataset for vehicle- infrastructure cooperative 3d object detection,","cited_arxiv_id":null,"evidence_quote":"Real-world vehicle-infrastructure dataset used for the DAIR-V2X experiments and delay robustness evaluation."},{"cited_title":"V2x-vit: Vehicle-to-everything cooperative perception with vision transformer,","cited_arxiv_id":null,"evidence_quote":"V2XSet dataset and the V2X-ViT baseline that CoDynTrust is compared against under asynchronous settings."},{"cited_title":"Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to- vehicle communication,","cited_arxiv_id":null,"evidence_quote":"OPV2V vehicle-to-vehicle dataset, specifically the Culver City Digital Twin test set used for generalization evaluation."}],"review_version":1}