{"id":"11dd4e65-fba9-414a-a5e5-08eaf6f080c6","arxiv_id":"2411.16457","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A trajectory prediction model built from a generic DDPM diffusion module and spatiotemporal attention reports strong RMSE gains that are not reproducible from the paper.","lead":"This paper proposes CDSTraj, a diffusion-plus-attention network for predicting vehicle trajectories, and claims it outperforms prior methods on three driving datasets. The claim is undercut by a formulation that feeds predicted future neighbor paths into the target prediction and by missing tables, error bars, and code.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central benchmark claim depends on Eq. (1)'s use of neighbor future trajectories Yi, whose source at inference is never specified; if they are ground truth, Table 1 is inflated by future-information leakage.","rationale":"The paper's central claim is explicitly quantitative: CDSTraj 'significantly outperforms existing state-of-the-art methods,' with Table 1 reporting 0.36 m at 1s and 2.85 m at 5s on NGSIM. For that claim to hold, the evaluation must measure genuine predictive performance, not performance obtained by conditioning on unobserved future information. Eq. (1) introduces exactly such an input: the target prediction Y0 is a function of neighboring agents' future trajectories Yi. Section 3 describes Yi as 'predicted future trajectories,' but it never specifies whether these come from the same model, from ground truth, or from an external predictor. This ambiguity is load-bearing because the reported gains could disappear entirely under reading (iii), could be artificially inflated under reading (i), or could constitute an incomparable joint-prediction paradigm under reading (ii). The methodology sections do not resolve the ambiguity: Eq. (5) conditions the denoiser only on X0 and Xi, and the decoder in Eq. (16) consumes the fused spatial-temporal state, with no mention of Yi. The paper therefore fails to establish that the evaluated system matches the formulation in Eq. (1). The reader's weakest_assumption identifies the same issue, and I agree that it is the most serious unresolved premise. The absence of HighD results, lack of error bars, and numerical inconsistencies in the text are additional signs of incomplete evidence, but the Eq. (1) ambiguity alone is sufficient to undermine the central claim. A controlled three-way ablation would settle the question directly. Since the concern is unaddressed and no code or data is released, the reader's REJECT verdict stands.","tokens_in":11201,"tokens_out":3510,"duration_ms":36231,"concrete_test":"Run a three-way ablation on NGSIM with the identical train/val split: (A) feed ground-truth neighbor future trajectories into Eq. (1); (B) feed the model's own predicted neighbor trajectories; (C) drop Yi and condition only on X0, Xi as Eq. (5) does. Compare 5s RMSE against Table 1's 2.85. If (A) matches and (C) degrades significantly, the headline improvement is attributable to future-information leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive weakness is the unresolved status of the neighbor-future input Yi in Eq. (1). The paper's headline results depend on this term being available at inference, but Section 3 never states its provenance, and Section 4.1's reverse diffusion (Eq. (5)) conditions only on X0 and Xi, not Yi. Three readings are left open: (i) Yi are ground-truth future trajectories, which would leak future information and inflate every horizon in Tables 1–3; (ii) Yi are the model's own predictions, making Phi a closed-loop joint predictor whose error accumulation is never analyzed; (iii) Yi is notationally vestigial, in which case the 'key innovation' claimed in Section 3 is not actually implemented. None of these readings is ablated or even discussed. Because the claim of state-of-the-art performance is quantitative, a result obtained under reading (i) would not support it; under reading (ii) the comparison with single-shot baselines would be unfair without retraining them in the same loop. This is not a stylistic objection: it determines whether the central empirical claim is valid at all.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CDSTraj, a trajectory prediction model for autonomous driving that combines a 'Characterized Diffusion Module' with a 'Spatial-Temporal Interaction Network.' The central idea, expressed in Eq. (1), is to predict a target agent's future trajectory from the agent's and neighbors' histories and from the predicted future trajectories of neighboring agents. The authors report experiments on NGSIM, HighD, and MoCAD and claim substantial improvements over existing methods; Table 1 reports NGSIM RMSE values of 0.36 m at 1 s and 2.85 m at 5 s. An ablation study over five components is also provided. The main unresolved issue is that the provenance of the neighbor-future term Yi is never specified, and the evidence for the headline claims is incomplete.","tokens_in":11450,"tokens_out":6823,"duration_ms":58510,"significance":"If the reported performance were reproducible, the 5 s NGSIM RMSE of 2.85 m would be a competitive result, and the combination of diffusion with spatial-temporal attention would be worth studying. The manuscript's strengths are the breadth of the baseline comparison in Table 1 and the effort to ablate the proposed components. However, the key formulation in Eq. (1) is ambiguous in a way that directly affects the benchmark numbers, no HighD result table is provided, and no error bars, seeds, or evaluation protocol details are given. As a consequence, the claimed state-of-the-art performance cannot be assessed or reproduced.","major_comments":[{"comment":"The central formulation Y0 = Phi(X0, Xi, Yi) makes the prediction of the target agent depend on the future trajectories Yi of neighboring agents, but the manuscript never states where Yi comes from at inference. In Section 4.1, the reverse diffusion in Eq. (5) conditions only on X0 and Xi, and the encoding and decoding stages in Sections 4.2-4.3 similarly do not consume Yi, so the equation is not implemented as written. If Yi are ground-truth future trajectories, the NGSIM and MoCAD results are inflated by information leakage; if Yi are the model's own outputs, the method is a closed-loop joint predictor whose error accumulation is not analyzed; if Yi is only notational, the stated key innovation is absent. The authors should specify the inference-time source of Yi and provide an ablation with and without this term.","section":"Section 3, Eq. (1); Section 4.1, Eq. (5)"},{"comment":"Section 5.3 reports large improvements on HighD (43%-70% for short-term predictions and 62%-78% for long-term predictions), but no HighD results table or numerical values appear anywhere in the manuscript. The MoCAD results are presented in Table 2 and the NGSIM results in Table 1, but the HighD claims are unverifiable. Either the HighD table should be added or the corresponding claims should be removed.","section":"Section 5.3"},{"comment":"No evaluation protocol is given: the manuscript does not state the train/validation/test split, the number of samples, the number of random seeds, or standard deviations for any reported RMSE value. In addition, the diffusion process produces K samples in Eq. (4), but the paper never explains how the K trajectories are converted into the scalar RMSE values in Tables 1-3, such as best-of-K, mean, or winner-take-all. Without this information, the comparison with single-shot baselines cannot be interpreted.","section":"Section 5, Tables 1-3"},{"comment":"The ablation study reports a single RMSE value per configuration without stating the prediction horizon, and the table has no error bars, so the reader cannot judge whether the differences between configurations are significant. The 'Confidence feature fusion' component, which is described as important in Section 5.4, is not defined anywhere in the methodology and is not listed among the ablated components in Table 3.","section":"Section 5.4, Table 3"},{"comment":"The quantitative claims in the abstract and text are not consistent with the table. The abstract says the model 'significantly outperforms' existing methods, but Table 1 shows BAT has lower RMSE at 1 s (0.23 vs 0.36) and at 2 s (0.81 vs 0.86). The text claims a 29% improvement over WSiP, but the table values (4.34 vs 2.85) imply a 34% improvement. The claims should be recomputed and restated to reflect the actual table entries.","section":"Abstract and Section 5.3, Table 1"}],"minor_comments":[{"comment":"Several citations are broken: Section 1 contains '[37; ?]' and Section 2's 'Xi et al. (2024)' has no corresponding numbered reference.","section":"Section 1 and Section 2"},{"comment":"The conclusion ends with the incomplete phrase 'the integra...' and should be completed.","section":"Section 6"},{"comment":"The four figures listed at the end of the manuscript (cdffusion1.png, exp111.png, head.png, network.png) are never referenced in the text and have no captions in the submitted text.","section":"Figures"},{"comment":"The NLL loss in Eq. (18) introduces alpha, rho_xy, and P_t without defining them or explaining how the NLL loss is combined with the MSE objective in the two-stage training procedure.","section":"Section 5.2, Eq. (18)"},{"comment":"The table caption states that bold and underlined values represent the best and second-best performance, but no underlining appears in the rendered table.","section":"Table 1"},{"comment":"The MoCAD dataset is described but no citation is provided for it, despite being one of the three evaluation benchmarks.","section":"Section 5.1"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an incomplete draft: at least one citation is a placeholder, the conclusion is cut off, and no code is provided. The unresolved status of Yi in Eq. (1) and the missing HighD results are the decisive issues; in the current form the empirical claims cannot be verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know before reading this: the paper's headline claim of state-of-the-art NGSIM results is not supported by the text as written. The core formulation in Eq. (1) makes the target prediction depend on the neighbor agents' predicted future trajectories Yi, but the method description never says where Yi comes from. The reverse diffusion update in Eq. (5) conditions only on historical states X0 and Xi. So either Yi is ground-truth future data (leakage), or it's the model's own output and we're in a closed loop that is never analyzed, or it's simply not used. The paper calls this the key innovation, so this is not a minor omission.\n\nWhat is genuinely there: the paper identifies a real problem—uncertainty in multi-agent trajectory prediction—and the writing is readable. The ablation study is the right idea, though without error bars or seeding information it doesn't establish much.\n\nThe soft spots are substantial. The claimed 29% improvement over WSiP is actually about 34% from the table values. The HighD improvements promised in Section 5.3 have no table. There are placeholder citations like [?] in the text, and the reference list is heavy with a single author cluster, some of which may not be cited properly. The diffusion update is exactly Ho et al.'s sampling step, so the novelty is limited.\n\nMy take: desk reject. The central empirical claim is unverifiable and likely inflated, and the core architectural novelty is not implemented as described. The authors could resubmit after clarifying the Yi issue, adding the missing HighD numbers, reporting error bars, and cleaning up the references. But as it stands, a referee would spend their time on a paper that doesn't hold together.\n\nRecommendation: pass on this one.","headline":"The paper's core formulation is not implemented as described; the benchmark claims are unsupported, so this is not ready for review.","tokens_in":11998,"tokens_out":4158,"would_cite":false,"duration_ms":36171,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A model coupling a characterized diffusion denoiser with a spatial-temporal attention network achieves state-of-the-art 5-second trajectory prediction on NGSIM, HighD, and MoCAD, with an RMSE of 2.85 meters on NGSIM.","keywords":["trajectory prediction","autonomous driving","diffusion models","spatial-temporal attention","uncertainty estimation","NGSIM","HighD","MoCAD"],"falsifier":"Run CDSTraj on NGSIM with $Y_i$ produced by the model's own decoder instead of ground-truth future frames and measure the 5-second RMSE; if the error rises from 2.85 m to roughly the 3.4-3.7 m range of the strongest baselines, the original result relied on oracle neighbor futures.","tokens_in":10947,"feed_emoji":"🚗","tokens_out":7208,"duration_ms":61631,"temperature":0.7,"pith_summary":"This paper proposes CDSTraj, a generative model for predicting the future paths of surrounding vehicles in autonomous driving. The model combines a Characterized Diffusion Module, which iteratively removes noise from candidate trajectories, with a Spatial-Temporal Interaction Network that attends to both the positions and the evolving behavior of neighboring agents. The authors claim this coupling addresses two shortcomings of prior work: focusing only on the target agent's uncertainty while ignoring the wider scene, and treating spatial interactions as static rather than time-varying. They report substantial gains, including a 5-second RMSE of 2.85 meters on NGSIM versus 3.40 meters for the next-best model, with similar improvements on HighD and MoCAD. If these results hold, the framework would offer an immediately usable accuracy boost for motion planning in dense and mixed-traffic settings.","feed_headline":"Cut 5-second trajectory error to 2.85 meters","feed_subtitle":"New model combines diffusion denoising with gated spatial-temporal attention to beat prior baselines on three road datasets.","key_machinery":"The central machinery is a two-part generative architecture. The Characterized Diffusion Module runs a forward noising process $C_\\delta = f_{\\text{diff}}(C_{\\delta-1})$ and a reverse denoising loop that conditions on historical states and adaptively scales updates via $\\alpha_\\delta$ and $\\bar{\\alpha}_\\delta$ in the rule $\\hat{C}_\\delta = \\frac{1}{\\sqrt{\\alpha_\\delta}}(\\hat{C}_{\\delta+1} - \\frac{1-\\alpha_\\delta}{\\sqrt{1-\\bar{\\alpha}_\\delta}} \\hat{\\epsilon}_\\delta) + \\sqrt{\\frac{1-\\alpha_\\delta}{\\alpha_\\delta}} z$. The Spatial-Temporal Interaction Network computes multi-head attention weights $\\omega = \\text{softmax}(QK^T/\\sqrt{d})$ over agents, then fuses spatial attention output $\\Upsilon$ with temporal recurrent states through a gate $S = H_a \\odot H_g$. An LSTM decoder turns the fused representation into predicted positions. The interaction between the two modules is the defining innovation: the diffusion module generates the neighbor futures $Y_i$ that the interaction network then uses to refine the target trajectory.","core_discovery":"On its own terms, the paper's central claim is that trajectory prediction accuracy improves when uncertainty is modeled by a diffusion process over the whole scene and when agent interactions are encoded jointly over space and time. Concretely, CDSTraj defines the target trajectory as $Y_0 = \\Phi(X_0, X_i, Y_i)$ for all neighbors $i$, meaning the prediction explicitly consumes the neighboring agents' future trajectories $Y_i$ alongside their histories. A forward diffusion process adds controlled noise to candidate trajectories and a reverse process denoises them using the historical context, with step-specific parameters $\\alpha_\\delta$ and $\\bar{\\alpha}_\\delta$ that adapt the update scale. The spatial-temporal module uses multi-head attention for pairwise interactions and a gated fusion $S = H_a \\odot H_g$ to combine the two dimensions. In experiments, the model reports RMSE values of 0.36, 0.86, 1.36, 2.02, and 2.85 meters at horizons 1 through 5 seconds on NGSIM, and similar improvements on HighD and MoCAD, with the gap over the strongest baselines growing at longer horizons.","pith_inferences":["The paper never specifies whether $Y_i$ in Eq. (1) come from ground-truth future labels, from the model's own predictions, or from an external predictor; if they are ground-truth, the reported benchmark gains would reflect oracle information, and a fair comparison would require ablating this input.","If $Y_i$ are model outputs, then CDSTraj is effectively a joint prediction model for all agents, and its inference cost and error accumulation across agents should be analyzed; the paper does not provide this analysis.","The uncertainty parameters ($\\sigma$ and $\\rho$) produced by the NLL stage could support risk-aware planning or collision-avoidance modules, an application the paper mentions in general terms but does not evaluate.","A natural extension is to replace the simple LSTM decoder with a trajectory-sampling head that outputs a full distribution, which would let the diffusion model's diversity be evaluated with metrics like minADE and miss rate rather than mean RMSE."],"forward_implications":["If the reported numbers hold, CDSTraj would be the most accurate published predictor on NGSIM, HighD, and MoCAD at horizons up to 5 seconds, with the largest margins in long-term predictions.","The two-stage training schedule (MSE followed by NLL) allows the model to output not only point trajectories but also per-step uncertainty parameters ($\\sigma^x_t$, $\\sigma^y_t$, $\\rho^{xy}_t$), which can quantify confidence in each predicted position.","The model's performance on MoCAD, which has left-hand traffic and varied urban conditions, suggests the architecture transfers across driving conventions without reconfiguration.","The ablation study attributes most of the gain to the characterized diffusion module and confidence feature fusion, since removing either raises NGSIM RMSE from 2.85 m to roughly 3.0-3.1 m."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model that the Characterized Diffusion Module builds on.","marker":"[25]"},{"why":"Foundational LSTM social pooling baseline that also motivates the recurrent temporal encoding.","marker":"[15]"},{"why":"Social GAN baseline, used as a comparison in all three dataset tables.","marker":"[27]"},{"why":"WSiP baseline, the prior state-of-the-art that CDSTraj claims to improve by 29% at 5 seconds.","marker":"[37]"},{"why":"STDAN baseline, improved by 22%, and source of the spatial-temporal dynamic attention idea.","marker":"[42]"},{"why":"Defines the HighD highway dataset used for evaluation.","marker":"[47]"}],"fun_headline_variants":["Diffusion + gated attention cuts 5s driving error to 2.85m","CDSTraj: spatial-temporal diffusion beats prior on three datasets","Denoising trajectories: attention + diffusion lowers 5s RMSE to 2.85m","Gated spatial-temporal attention with diffusion improves driving forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's defining equation $Y_0 = \\Phi(X_0, X_i, Y_i)$ requires the future trajectories of neighboring agents, and the paper never states whether those come from ground truth, from the model itself, or from an external predictor, so the reported accuracy gains may depend on information that is not available at inference time.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion + gated attention cuts 5s driving error to 2.85m","CDSTraj: spatial-temporal diffusion beats prior on three datasets","Denoising trajectories: attention + diffusion lowers 5s RMSE to 2.85m","Gated spatial-temporal attention with diffusion improves driving forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1536,"prompt_tokens":884,"completion_tokens":652,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":569}},"tokens_in":500,"tokens_out":652,"duration_ms":7512,"temperature":1.0,"reasoning_tokens":569,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:06:29.389907+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CDSTraj on NGSIM with $Y_i$ produced by the model's own decoder instead of ground-truth future frames and measure the 5-second RMSE; if the error rises from 2.85 m to roughly the 3.4-3.7 m range of the strongest baselines, the original result relied on oracle neighbor futures.","supporting_citations":[{"cited_title":"Social gan: Socially acceptable trajectories with generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Social GAN baseline, used as a comparison in all three dataset tables."},{"cited_title":"Social lstm: Human tra- jectory prediction in crowded spaces,","cited_arxiv_id":null,"evidence_quote":"Foundational LSTM social pooling baseline that also motivates the recurrent temporal encoding."},{"cited_title":"Wsip: Wave superposition inspired pooling for dynamic interactions-aware trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"WSiP baseline, the prior state-of-the-art that CDSTraj claims to improve by 29% at 5 seconds."},{"cited_title":"Intention-aware vehicle trajectory prediction based on spatial-temporal dynamic attention network for internet of vehicles,","cited_arxiv_id":null,"evidence_quote":"STDAN baseline, improved by 22%, and source of the spatial-temporal dynamic attention idea."},{"cited_title":"The highd dataset: A drone dataset of naturalistic ve- hicle trajectories on german highways for validation of highly automated driving systems,","cited_arxiv_id":null,"evidence_quote":"Defines the HighD highway dataset used for evaluation."}],"review_version":1}