{"id":"dbf000b2-5f55-426d-b0b1-f1627861d61f","arxiv_id":"2505.06856","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A causal inference framework with backdoor adjustment and counterfactual analysis reports accuracy gains on five driving datasets, but the causal derivation is weakly grounded.","lead":"This paper presents a trajectory prediction model that uses causal reasoning to avoid spurious patterns in driving data. If its accuracy and robustness claims hold up, it could make autonomous driving systems safer in unseen scenarios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal-identification claim is unsupported: Eq. (1) uses uniform P(s_i)=1/n without evidence that diffusion samples approximate P(S), Eq. (2) is not a Pearl counterfactual, and S still enters through BEV; empirical gains are not attributable to causal adjustment.","rationale":"The paper's stated contribution is causal inference for trajectory prediction, and the formal derivation in Eqs. (1)-(2) is the only support for that contribution. The reader's weakest assumption correctly identifies that the diffusion samples and uniform weights may not implement backdoor adjustment. My analysis sharpens this: even if the samples were perfect draws from P(S), Eq. (2) is not a counterfactual estimand, and the BEV encoder still feeds the raw road layout S into the prediction path, so the backdoor adjustment is not cleanly applied. This means the causal mechanism cannot be verified from the paper as written. The empirical superiority claim in Tables 1-3 is separable and may be reproducible, but the paper's central causal narrative would not follow. The proposed synthetic-data test is a concrete way to check whether the implemented Eqs. (1)-(2) recover a known ground-truth interventional distribution; if they do not, the causal interpretation of the reported gains should be withdrawn or substantially weakened. Because the reader already issued a conditional verdict, my read does not change that verdict, but it confirms that the condition is substantive.","tokens_in":13118,"tokens_out":7008,"duration_ms":82263,"concrete_test":"Build a synthetic dataset with known structural equations and a non-uniform discrete confounder S, e.g., S in {1,2,3} with P(S)=(0.6,0.3,0.1), where S influences X and Y but T does not cause X, with X -> Y and T -> Y. Train the proposed model with its diffusion-based backdoor adjustment and evaluate the estimator in Eq. (1) with P(s_i)=1/n. Compute the exact backdoor-adjusted interventional distribution P(Y|do(X=x),T=t) from the known structural equations and compare the two on held-out (x,t) grids using a distributional metric such as KL divergence or maximum mean discrepancy. If the difference exceeds the model's own seed variability, the implemented backdoor adjustment is not estimating the causal quantity claimed, and the causal interpretation of Tables 1 and 4 does not follow.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim rests on Eqs. (1)-(2) in Section 3.2 and the diffusion-based backdoor adjustment in Section 3.3. For Eq. (1) to be a valid backdoor adjustment, g_theta must estimate P(Y|X,T,S=s_i), the s_i must be drawn from the true P(S), and P(s_i) must be the corresponding probability mass. None of these conditions is established. The paper asserts P(s_i)=1/n on a 'maximum entropy' principle, which is only correct for a uniform prior, and the diffusion-generated spatial tokens are not shown to approximate any meaningful stratum distribution over S. Eq. (2) is also not Pearl counterfactual inference: zeroing the historical trajectory in the progressive fusion module does not implement do(X=X_c) or an abduction step, and 'Y = Y_tilde - Y_tilde_c' is not a standard causal estimand. Furthermore, the causal graph labels T a confounder although no edge T -> X is drawn, and the actual road layout S is still injected through the BEV encoder in Eqs. (5)-(7). This means the purported severing of S -> X and S -> T is not actually achieved. The empirical accuracy numbers may reproduce, but they cannot be attributed to the causal machinery as described.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a trajectory prediction framework for autonomous driving that combines spatial, BEV, and temporal encoders with a diffusion-based backdoor adjustment module, counterfactual analysis via zeroed historical trajectories, and a cross-modal progressive fusion decoder. The authors claim that this causal inference paradigm removes spurious correlations, improves robustness and generalization, and achieves state-of-the-art accuracy on five real-world datasets (ApolloScape, nuScenes, NGSIM, HighD, MoCAD), while also reporting efficiency and plug-and-play validation results.","tokens_in":13518,"tokens_out":3288,"duration_ms":33062,"significance":"If the causal identification claims were valid, the paper would offer a principled way to reduce confounder bias in trajectory prediction and would strengthen the case for causal methods in autonomous driving. The empirical contribution is substantial: evaluation across five datasets with multiple metrics, robustness to noise and missing frames, a domain generalization study, a parameter-efficient real-time model, and a plug-and-play integration into an existing baseline. However, the causal claims are not established by the derivations or experiments as written; the central contribution is currently an architecture with causal labels rather than a demonstrated causal estimator.","major_comments":[{"comment":"The backdoor adjustment formula in Eq. (1) is not derived from the stated causal graph. For Eq. (1) to be a valid backdoor adjustment, g_theta must estimate P(Y|X,T,S=s_i), the s_i must be sampled from the true confounder distribution P(S), and P(s_i) must be the corresponding probability masses. The paper instead sets P(s_i)=1/n on a 'maximum entropy' principle without justifying uniform weights over diffusion-generated tokens, and it does not show that the diffusion samples approximate any meaningful stratum distribution over S. As written, Eq. (1) is a model definition, not an identifiable causal estimand.","section":"Section 3.2, Eq. (1)"},{"comment":"The counterfactual analysis in Eq. (2) is not Pearl counterfactual inference. Setting X_c to zero vectors in the progressive fusion module does not specify an abduction step, an intervention do(X=X_c), or a prediction step on a modified structural model. Consequently, the composite output Y = Y_tilde - Y_tilde_c is not a standard causal estimand, and no identification argument is provided for this subtraction. The counterfactual branch is a heuristic architectural choice rather than a causal operation.","section":"Section 3.2, Eqs. (2) and (3)"},{"comment":"The causal graph labels T as a confounder, but Figure 1(c) does not draw an edge T -> X. A confounder of the effect of X on Y requires a backdoor path with an arrow into X. Either the graph is mis-specified or T is not a confounder, which undermines the motivation for using counterfactual subtraction to 'eliminate the confounding effects of temporal agent data T.' Additionally, the claim that the diffusion-based backdoor adjustment severs S -> X and S -> T is contradicted by Eqs. (5)-(7), where the BEV encoder still injects spatial information S through B^h and the spatial token S^{h,i} is used directly in attention.","section":"Section 3.2, Figure 1(c)"},{"comment":"The ablation study does not isolate the causal adjustment. Method D removes all causal modules and simultaneously replaces the decoder with a simple GRU, so the performance drop could be due to the diffusion augmentation, the attention modules, or the decoder architecture rather than the causal operations. Moreover, Tables 1-4 and 7 report no error bars, confidence intervals, or repeated-trial statistics, so the claimed improvements (for example, a 1.84% WSADE gain on ApolloScape) are not shown to be statistically reliable.","section":"Section 4.3, Tables 6 and 7"}],"minor_comments":[{"comment":"The heading 'Casual Inference' should be 'Causal Inference'.","section":"Section 3.2"},{"comment":"There is an inconsistent spacing in 'A Vs' and 'A V'; please use a consistent notation such as 'AVs' or 'autonomous vehicles'.","section":"Section 1"},{"comment":"The caption states 'Bold and underlined values represent the best and second-best performance,' but the table does not visibly use underlining.","section":"Table 1 caption"},{"comment":"The same BAT baseline appears with different reference years ('Liao et al., 2024c' and 'Liao et al., 2024d') in the HighD and MoCAD rows; please verify the citations.","section":"Table 3"},{"comment":"The definition of L0 is described as 'varies by dataset,' but the precise form for each dataset is not given; please specify the exact formula used for each metric.","section":"Section 3.4, Eq. (10)"},{"comment":"The subfigures in Figure 3 lack axis labels and error bars; please add these so the minADE5 values and variability can be assessed.","section":"Figure 3"},{"comment":"There is a typo in 'these results validate the prediction accurancy of our model'; 'accurancy' should be 'accuracy'.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical breadth is a strength, but the causal inference framing is currently unsupported by the derivations and experiments. The authors should either provide a rigorous identification argument for Eqs. (1)-(3) and the counterfactual subtraction, or substantially reframe the contribution as a non-causal architectural innovation. Without one of these changes, the central claim of the paper cannot be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The empirical architecture is a solid, competitive trajectory predictor, but the paper's central claim—that it performs backdoor adjustment and counterfactual inference—is not supported by the equations as written. The model may be worth a serious referee, but the causal section needs to be rewritten or downgraded.\n\nWhat is genuinely new: the combination of diffusion-generated spatial tokens, progressive fusion, and dual-scale fusion is not present in the cited prior work. The evaluation spans five real datasets, includes ablations, robustness under noise and dropped frames, parameter counts, latency, and a plug-and-play variant on PGP. That is more evidence than many papers in this area provide. If the numbers reproduce, the architecture itself is a useful contribution to the trajectory-prediction subfield.\n\nThe soft spots are mostly concentrated in the causal derivation, and they are real. Eq. (1) assumes P(s_i)=1/n on a 'maximum entropy' basis, but no argument shows the diffusion samples approximate the true distribution over road layouts S. Eq. (2) is not a Pearl counterfactual: zeroing the history and rerunning the fusion module does not implement do(X=X_c) or an abduction step, and Y=Y_tilde−Y_tilde_c is a heuristic subtraction, not an identifiable estimand. The causal graph calls T a confounder without drawing an edge T→X, and S is still fed through the BEV encoder in Eqs. (5)–(7), so the claimed severing of S→X and S→T is not actually achieved. These are load-bearing problems because the causal story is the paper's main selling point.\n\nThe empirical section is better but not airtight. No error bars, no code release, and the MoCAD dataset comes from the authors' own prior work. The 'domain generalization' split uses only four subsets of nuScenes and a KS test; it is suggestive, not conclusive. The robustness test with injected noise is reasonable, but the causal attribution of the gains is unproven.\n\nBottom line: if I were reviewing, I would ask for a major revision that either replaces the causal framing with a 'confounder-robust architecture' claim or actually justifies the backdoor adjustment with a clear distributional argument and a proper counterfactual baseline. This is not a desk reject: the empirical results are potentially useful and the architecture is fresh. It deserves a serious referee, but not unqualified acceptance.\n\nRecommendation: send to peer review, with the expectation that the causal claims will be either considerably weakened or properly derived.","headline":"A plausible trajectory predictor with competitive numbers, but the causal inference story is not derived—send to review, don't take the causal claims at face value.","tokens_in":13934,"tokens_out":3570,"would_cite":false,"duration_ms":37427,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims to beat prior trajectory predictors on five driving datasets by treating the road map and nearby agents as confounders and adjusting for them causally.","keywords":["trajectory prediction","causal inference","backdoor adjustment","counterfactual analysis","autonomous driving","diffusion model","domain generalization","spurious correlation"],"falsifier":"Keep the architecture identical but replace the diffusion-based backdoor module with the single observed road layout, so $n=1$ and no stratification occurs: if error metrics barely move, the adjustment term is not carrying the claimed benefit. The sharper test is synthetic: generate traffic scenes with a known confounder, train the causal model and a purely correlational baseline on them, and check whether the causal model recovers the true intervention effect, for instance whether it learns to predict a stop at a crosswalk when pedestrians are present, purely from the known causal structure.","tokens_in":12978,"feed_emoji":"🚗","tokens_out":13497,"duration_ms":106195,"temperature":0.7,"pith_summary":"This paper tries to establish that causal inference, not just pattern matching, is what makes trajectory prediction generalize across driving scenes. The authors build a causal graph in which the target vehicle's past motion and the surrounding agents cause the future trajectory, while the road layout confounds everything, creating backdoor paths that a purely correlational model would learn as spurious shortcuts. They implement backdoor adjustment by generating $n$ alternative road layouts with a diffusion model and averaging predictions over them with equal weight $1/n$, then subtract a counterfactual prediction obtained by zeroing out the target's history to remove the remaining confounding influence of other agents. On ApolloScape, nuScenes, NGSIM, HighD, and MoCAD the resulting model reports lower error metrics than the compared baselines, stays accurate under added noise and dropped frames, and keeps its advantage when the causal module is grafted onto another predictor. If the causal attribution is correct, this is a route to predictors that trust the actual cause of a maneuver, such as a pedestrian entering a crosswalk, rather than a correlation learned from the training distribution.","feed_headline":"Causal logic beats pattern-matching in driving trajectory prediction","feed_subtitle":"Treating the road map and other drivers as confounders, the model drops spurious correlations and wins on five datasets.","key_machinery":"The load-bearing object is the causal graph $X \\to Y$, $T \\to Y$ with $S \\to X$ and $S \\to T$ creating backdoor paths, operationalized as a pair of adjustment equations: backdoor adjustment $\\tilde{Y} = \\sum_{i=1}^{n} g_\\theta(X, S=s_i, T)P(s_i)$ with uniform weight $P(s_i)=1/n$, where the $n$ road layouts $s_i$ are produced by a diffusion model that noisifies and regenerates the encoded map token; and counterfactual adjustment $\\tilde{Y}_c = \\sum_{i=1}^{n} g_\\theta(\\mathrm{do}(X=X_c), S=s_i, T)P(s_i)$, where $X_c$ zeroes out the target agent's history. Subtracting the counterfactual from the factual, $Y = \\tilde{Y} - \\tilde{Y}_c$, is meant to cancel the non-causal contribution of the temporal and spatial context. Around this identity sit the spatial, BEV, and temporal encoders that tokenize each input, a targeted multi-view attention module fusing the tokens, and a cross-modal progressive-fusion decoder that refines an anchor query over several stages, alongside a dual-scale CNN branch, feeding a causal decoder that emits multimodal trajectory hypotheses. The diffusion-based sampling is what makes the stratification in the backdoor sum tractable: it generates the alternative road layouts that the adjustment averages over.","core_discovery":"The paper's central claim is that decomposing the driving environment into a spatial component $S$ and a temporal component $T$, and then intervening on both through causal machinery, removes the spurious correlations that bind ordinary predictors to their training scenes. The model computes $\\tilde{Y} = \\sum_{i=1}^{n} g_\\theta(X, S=s_i, T)P(s_i)$ with $P(s_i)=1/n$, which is backdoor adjustment over diffusion-generated road layouts, and the counterfactual $\\tilde{Y}_c = \\sum_{i=1}^{n} g_\\theta(\\mathrm{do}(X=X_c), S=s_i, T)P(s_i)$, in which the target's history is replaced by a counterfactual value; the final prediction is the difference $Y = \\tilde{Y} - \\tilde{Y}_c$. Trained in two stages, with diffusion loss first and then the full model, the predictor reports lower WSADE and WSFDE on ApolloScape, minADE and FDE on nuScenes, and RMSE on NGSIM, HighD, and MoCAD than the compared baselines, keeps 0.28M parameters with 57ms inference on nuScenes, and stays ahead under added noise and frame dropout. The causal module also plugs into the PGP baseline and improves its turning predictions, while ablating the causal components degrades performance the most.","pith_inferences":["The uniform prior $P(s_i)=1/n$ is the most stress-testable design choice: a prior learned from road-layout statistics, or one conditioned on traffic rules, could sharpen the backdoor sum if the diffusion samples are informative about the true confounder distribution.","The same backdoor-plus-counterfactual subtraction could transfer to other spatiotemporal forecasting problems, such as pedestrian intent or vessel and drone motion, wherever a static scene layout confounds an agent's observed behavior.","A direct test of whether the causal story rather than the extra parameters carries the gain: keep the architecture fixed and set $n=1$ so the backdoor sum collapses to the single observed layout; if metrics barely move, the adjustment is not the active ingredient.","The combination rule $Y = \\tilde{Y} - \\tilde{Y}_c$ treats the causal effect as additive; whether a ratio or a learned gating of factual and counterfactual terms would serve better is a variant the paper does not test."],"forward_implications":["If the causal decomposition is correct, predictors can ignore correlations that do not survive intervention, so accuracy on rare and unseen scenes, including crosswalks, intersections, and unfamiliar road layouts, should improve without new labeled data.","Because the causal module also improves the PGP baseline when plugged in, existing predictors could adopt the module without a full redesign, making the method an upgrade rather than a replacement.","The robustness results imply that under sensor noise and missing frames, modeled as curvature-scaled Gaussian noise and random frame dropout, prediction quality degrades more slowly than for the compared baselines, which matters for real perception stacks.","The efficiency figures of 0.28M parameters and 57ms per sample on nuScenes imply the causal machinery does not add a latency barrier to real-time use."],"supporting_citations":[{"why":"Supplies the backdoor adjustment and do-calculus formalism that Eqs. (1) and (2) instantiate.","marker":"[Pearl, 2009]"},{"why":"Prior backdoor adjustment for trajectory prediction; the source of the uniform weight P(s_i)=1/n maximum-entropy choice used in Eq. (1).","marker":"[Ge et al., 2023]"},{"why":"Prior counterfactual analysis for trajectory prediction that the paper's counterfactual branch extends.","marker":"[Chen et al., 2021]"},{"why":"Supplies the nuScenes dataset used for the main benchmark, robustness, and efficiency tables.","marker":"[Caesar et al., 2020]"},{"why":"Supplies the ApolloScape dataset and the WSADE/WSFDE evaluation used in Table 1.","marker":"[Huang et al., 2018]"},{"why":"Supplies the HighD highway dataset used in Table 3.","marker":"[Krajewski et al., 2018]"},{"why":"Source of the NGSIM dataset and of the CS-LSTM baseline in Table 3.","marker":"[Deo and Trivedi, 2018]"},{"why":"The BAT baseline on HighD and MoCAD, and the cited source of the MoCAD dataset and evaluation conventions.","marker":"[Liao et al., 2024c]"},{"why":"PGP, the baseline the causal module is plugged into in Section 4.4 and compared against on nuScenes.","marker":"[Deo et al., 2022]"},{"why":"LAformer, the strongest nuScenes baseline reported in Table 2 that the model must beat.","marker":"[Liu et al., 2024]"}],"fun_headline_variants":["Causal logic outruns pattern-matching in trajectory prediction","Backdoor adjustment wins on five driving datasets","Counterfactual reasoning beats spurious correlations in driving","Causal model drives autonomous trajectory prediction gains","Drop causal shortcuts: new model wins on five roads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the road layout $S$ really is a confounder of the target's future trajectory and that averaging over diffusion-generated layouts, each weighted equally at $1/n$, faithfully implements backdoor adjustment; if the generated layouts do not track the true distribution of road layouts, or if an unmodeled confounder such as traffic-signal phase or driver identity is the real driver of behavior, the causal claims do not follow even if the accuracy numbers reproduce.","fun_headline_variants_meta":{"raw":{"variants":["Causal logic outruns pattern-matching in trajectory prediction","Backdoor adjustment wins on five driving datasets","Counterfactual reasoning beats spurious correlations in driving","Causal model drives autonomous trajectory prediction gains","Drop causal shortcuts: new model wins on five roads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1625,"prompt_tokens":998,"completion_tokens":627,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":553}},"tokens_in":614,"tokens_out":627,"duration_ms":6609,"temperature":1.0,"reasoning_tokens":553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:30:49.283694+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep the architecture identical but replace the diffusion-based backdoor module with the single observed road layout, so $n=1$ and no stratification occurs: if error metrics barely move, the adjustment term is not carrying the claimed benefit. The sharper test is synthetic: generate traffic scenes with a known confounder, train the causal model and a purely correlational baseline on them, and check whether the causal model recovers the true intervention effect, for instance whether it learns to predict a stop at a crosswalk when pedestrians are present, purely from the known causal structure.","supporting_citations":[{"cited_title":"Causality","cited_arxiv_id":null,"evidence_quote":"Supplies the backdoor adjustment and do-calculus formalism that Eqs. (1) and (2) instantiate."},{"cited_title":"Human trajectory prediction via counter- factual analysis","cited_arxiv_id":null,"evidence_quote":"Prior counterfactual analysis for trajectory prediction that the paper's counterfactual branch extends."},{"cited_title":"nuscenes: A multimodal dataset for autonomous driving","cited_arxiv_id":null,"evidence_quote":"Supplies the nuScenes dataset used for the main benchmark, robustness, and efficiency tables."},{"cited_title":"The apolloscape dataset for au- tonomous driving","cited_arxiv_id":null,"evidence_quote":"Supplies the ApolloScape dataset and the WSADE/WSFDE evaluation used in Table 1."},{"cited_title":"The highd dataset: A drone dataset of naturalistic vehicle trajectories on ger- man highways for validation of highly automated driving systems","cited_arxiv_id":null,"evidence_quote":"Supplies the HighD highway dataset used in Table 3."},{"cited_title":"Convolutional social pooling for vehicle trajec- tory prediction","cited_arxiv_id":null,"evidence_quote":"Source of the NGSIM dataset and of the CS-LSTM baseline in Table 3."},{"cited_title":"Multimodal trajectory prediction conditioned on lane-graph traversals","cited_arxiv_id":null,"evidence_quote":"PGP, the baseline the causal module is plugged into in Section 4.4 and compared against on nuScenes."},{"cited_title":"Laformer: Trajectory pre- diction for autonomous driving with lane-aware scene con- straints","cited_arxiv_id":null,"evidence_quote":"LAformer, the strongest nuScenes baseline reported in Table 2 that the model must beat."}],"review_version":1}