{"id":"3d4810e5-f407-4607-b878-8671fe5290eb","arxiv_id":"2411.12313","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"C2INet couples a variational causal intervention with a continually updated prior queue to reduce catastrophic forgetting and improve multi-agent trajectory prediction.","lead":"C2INet is a continual learning method for multi-agent trajectory prediction that models environmental confounders with a variational encoder and stores scene-specific priors in a memory queue. The authors report improved average and final displacement errors over several baselines on ETH-UCY, a synthetic dataset, and the Stanford Drone Dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq.3 silently drops the latent representation Z from the variational objective, so the optimized loss is not a valid ELBO for the claimed backdoor adjustment; the causal intervention claim is therefore unsupported.","rationale":"The paper's central claim is that C2INet 'effectively mitigating confounding factors' through continual causal intervention. The most load-bearing step is the variational derivation of the intervention objective in Sec.2.3. I examined Eq.3 line by line. The move from the second to the third displayed equation drops the expectation over P(Z) and removes the P(Z|X,C) factor from the logarithm. No justification is given; Z is not marginalized by a deterministic function, since Eq.4 introduces a stochastic Gaussian mixture for P(Z|X,C). With Z still latent, the correct ELBO must include a variational term for Z (e.g., E_{Q(C|X)} KL(P(Z|X,C) || R(Z|X,Y,C))), plus the corresponding reconstruction term. The paper's final objective only regularizes C, so the model is a conditional VAE on C with Z treated as an intermediate deterministic feature. This invalidates the causal interpretation: even if the causal graph in Fig.1 is correct, optimizing Eq.3 does not compute P(Y|do(X)). As a result, the claimed mechanism of causal debiasing is unsupported, and the performance gains over baselines could simply be due to the memory queue, the symmetric KL penalty, and task-specific priors, which are standard continual-learning ingredients. The reader's weakest_assumption correctly notes that the causal identification is untested; my concern is stronger: the derivation itself is invalid, so the method does not implement the claimed backdoor adjustment regardless of identifiability. Other issues (Eq.9's ill-defined KL to a non-distribution, missing error bars/statistics, no code) reinforce the rejection, but this is the single load-bearing flaw. A direct re-derivation of Eq.3 with Z marginalized would settle the matter.","tokens_in":19241,"tokens_out":7700,"duration_ms":64801,"concrete_test":"Re-derive Eq.3 from the joint distribution P(Y,Z,C|X) by properly marginalizing Z (i.e., integrating E_{P(Z|X,C)}) and applying Jensen's inequality; verify whether the correct ELBO contains an additional term such as E_{Q(C|X)} KL(P(Z|X,C) || R(Z|X,Y,C)) that is absent from Eq.3. If the re-derived bound differs from Eq.3, the optimized objective is not the causal interventional bound and the causal claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that C2INet performs causal intervention via backdoor adjustment rests on the derivation of Eq.3. In that derivation, the expectation over the latent trajectory representation Z is silently discarded: the integrand changes from E_{P(Z),Q(C|X)} log [P(Y|X,Z,C) P(Z|X,C) P(C) Q / (P(C|X,Y,Z) Q)] to E_{Q(C|X)} log [P(Y|X,Z,C) P(C) Q / (P(C|X,Y,Z) Q)], removing both the P(Z|X,C) factor and the integration over Z. Since Z is stochastic (Eq.4 defines P(Z|X,C) as a Gaussian mixture), the resulting objective E_Q[log P(Y|X,Z,C)] - KL(Q(C|X)||P(C)) is not a valid ELBO for the joint model P(Y,Z,C|X). Consequently, optimizing Eq.3 does not implement P(Y|do(X)); the method reduces to a conditional VAE with a learned prior, and the reported gains could arise from the memory queue and KL regularization rather than causal debiasing. This is load-bearing because the paper's distinctive contribution is the causal intervention mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes C2INet, a continual-learning framework for multi-agent trajectory prediction. The authors model environmental context as a confounder C, use backdoor adjustment to define P(Y|do(X)), and approximate the resulting interventional distribution with a variational objective in which an encoder Q(C|X) estimates C and a memory queue stores scenario-specific priors. They introduce online and offline modes, a min-max training scheme with pruning, and report experiments on ETH-UCY, a synthetic dataset, and SDD, comparing against several continual-learning and causal baselines. The paper's central claim is that C2INet consistently improves trajectory prediction while mitigating confounders and catastrophic forgetting.","tokens_in":19635,"tokens_out":6132,"duration_ms":73054,"significance":"If the causal-intervention formalism were sound, the paper would address a real and currently relevant problem: continual trajectory prediction under environmental shift. The work has practical strengths: it embeds the method in two backbones, evaluates on multiple datasets, includes an ablation study, and reports five-seed averages. However, the theoretical core is not sound as written. The derivation of Eq. (3) loses the latent representation Z and an unknown posterior term, so the optimized objective is not a valid ELBO for the backdoor-adjusted interventional distribution. The continual-prior update in Eq. (9) minimizes a quantity that is not a KL divergence, and the prior in Eq. (8) is constructed from the same posterior it is meant to constrain. Since the manuscript's distinctive contribution is the causal-intervention mechanism, these issues undermine the central claim and reduce the contribution to a regularized continual VAE with a memory queue.","major_comments":[{"comment":"The variational derivation is invalid. In the transition between the third and fourth lines of Eq. (3), the expectation over P(Z) is silently replaced by an expectation over Q(C|X) alone, while the factor P(Z|X,C) is dropped; the integration over Z disappears. The paper also drops the term KL(Q(C|X)||P(C|X,Y,Z)) and treats the remaining expression as a lower bound, but dropping a KL term without replacing it by a bound is not a valid ELBO manipulation. Consequently, optimizing Eq. (3) does not implement P(Y|do(X)) under the stated graphical model; it is a conditional VAE objective with an added regularizer. This is load-bearing because the paper's causal-intervention claim rests on Eq. (3).","section":"Sec. 2.3, Eq. (3)"},{"comment":"The first-order expansion of the KL divergence in Eq. (7) is asserted without conditions on α_{K-1}, and the resulting derivative term, written as −α_{K-1}(P̂_K(C) log(Q_i(C|X)/M_{≤K−1}(C)) − 1), is not a legitimate functional derivative of the KL expression. More seriously, Eq. (9) minimizes KL[P̂_K(C) || Σ_i Q_i(C|X)/M_{≤K−1}(C) − 1], but the second argument is not a probability distribution: it can be negative and does not integrate to one. Therefore the quantity inside the proposed KL is not a valid density, and the optimization problem is not well-defined as a variational inference step. The continual-prior algorithm needs to be redefined with a proper divergence or a correctly normalized surrogate.","section":"Sec. 3, Eqs. (7) and (9)"},{"comment":"The prior P(C) is defined as an average of the model's own posterior Q(C|U_i) over trainable pseudo-features U_i. These pseudo-features are then optimized under KL constraints that involve the same posterior family. The prior is therefore not an independent environmental distribution, which is required for the backdoor adjustment in Eq. (1) to be interpreted causally. As a result, the claimed debiasing mechanism collapses into a learned prior in a regularized VAE, and the causal identification assumptions in Sec. 2.2 are untested. The authors need to either provide a causal ground-truth or sensitivity analysis, or explicitly reframe the method as a continual prior-learning method without causal claims.","section":"Sec. 3, Eq. (8)"},{"comment":"The empirical evaluation does not establish that the reported gains come from the causal mechanism or even from the proposed continual prior, because several critical hyperparameters are tuned on the same tasks used for evaluation. The queue capacity γ is selected via Fig. 8 on ETH-UCY, the task-order sensitivity is measured on the same dataset in Fig. 9, and the weights α are optimized online during training. Since the objective already reduces to a CVAE-like loss, the performance improvement over baselines could be explained by memory replay and KL regularization. The authors should report a validation/tuning protocol separate from test tasks, include variance estimates, and compare against a non-causal variational continual-learning baseline with matched capacity.","section":"Sec. 5 and Appendix A.3.4/A.3.5"}],"minor_comments":[{"comment":"The sentence 'According to the Markov property, P(Y|X,Z,C) = P(Y|X,Z)P(Z|X,C)' is not a standard use of the Markov property in this graph; the equality mixes a conditional distribution over Y with a density over Z and should be written with the appropriate factorization and integrations, if it is intended at all.","section":"Sec. 2.3, text after Eq. (3)"},{"comment":"The notation is inconsistent: Fig. 1 labels a 'Trajectory Prior P(X)' and a 'Distribution Mixture P(Z|X,C)', but the paper's equations use P(C) as the prior and define P(Z|X,C) as a mixture of Gaussians; the caption and the equations should be aligned.","section":"Fig. 1 and Sec. 2.3"},{"comment":"The table columns and the role of the leading colon in task names are not clearly explained; the repeated baseline names in the header make the table difficult to read, and the absence of standard deviations makes it hard to assess whether the reported improvements are statistically significant.","section":"Table 1 caption and formatting"},{"comment":"The condition `if j mod L/⌊2γ⌋ = 0` is ambiguously formatted and could be undefined for some values of γ; please rewrite the modulo condition unambiguously.","section":"Algorithm 1, line 12"}],"recommendation":"reject","confidential_remarks":"The manuscript has a serious gap between its causal framing and its actual derivation. The central equation (Eq. 3) is not a valid ELBO for the backdoor-adjusted interventional distribution, and the continual-prior update (Eq. 9) is not a valid KL minimization. These are not merely presentation issues; they are load-bearing for the claim of causal intervention. The empirical study is extensive, but the selection of hyperparameters on the same test datasets and the absence of a non-causal matched baseline leave the measured improvements ambiguous. If the authors resubmit, I would want the theoretical derivation replaced with a correct variational lower bound, the continual-prior objective redefined in a well-posed form, and a validation protocol that separates tuning from testing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nC2INet is a solid engineering idea in search of a trustworthy theory. What's genuinely new is the package: a memory queue of learned pseudo-feature priors, online and offline update modes, a pruning rule, and plug-and-play integration with trajectory encoders like STGAT and SocialSTGCNN. The authors clearly know the relevant literature—VampPrior, BOO-VAE, Pearl's backdoor adjustment, and the continual-learning baselines—and the citations land where they should. The three-dataset evaluation, with separate per-task results, ablations, and a task-order sensitivity check, is more than most papers in this space bother to do.\n\nThe math, however, does not carry the load. Eq.3 is the centerpiece: between the second and third lines the expectation over the latent trajectory representation Z silently vanishes, along with the P(Z|X,C) factor. Z is stochastic in Eq.4, so the objective that gets optimized is not a lower bound on the claimed intervention distribution P(Y|do(X)). It is a conditional VAE with a learned prior. That means the distinctive claim of the paper—that C2INet performs debiasing through causal intervention—is unsupported by the derivation. The stress-test note is right, and this is the correct reason to be skeptical.\n\nTwo smaller but real issues: Eq.9 minimizes KL against (sum_i Q_i/M - 1), which is not a valid probability distribution, and Eq.7 uses a first-order expansion without checking the remainder. The second might be patchable; the first is just wrong as written. The empirical section also has weaknesses: no error bars or significance tests despite five seeds, no code release or implementation details, and the identifying assumption that a single latent C captures all confounders is never probed.\n\nWho should read this? People building continual trajectory predictors can take practical ideas from the queue and pruning design. People looking for a causal intervention method should not take Eq.3 at face value.\n\nMy recommendation is to send it to review, not desk-reject. The pipeline is plausible, and the empirical gains may survive even after the causal language is scaled back. A good referee should ask the authors to fix the derivation or reframe the method as a continual VAE with learned priors and memory replay. That is a substantial but manageable revision, and the paper would be more honest and still useful afterward.","headline":"Useful continual-learning machinery, but the causal claim does not survive contact with the paper's own equations; the method is closest to a VAE with learned priors plus memory replay.","tokens_in":20113,"tokens_out":3919,"would_cite":false,"duration_ms":40432,"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":"This paper claims that a continual causal-intervention module, which estimates scene-specific confounders and stores their priors in a bounded memory queue, reduces catastrophic forgetting and beats prior continual and causal baselines in…","keywords":["trajectory prediction","continual learning","causal intervention","backdoor adjustment","variational inference","catastrophic forgetting","multi-agent motion forecasting","memory replay"],"falsifier":"Train the model on a synthetic dataset where the ground-truth confound is known, such as a rule that shifts all right turns in one scene, then test whether the learned $Q(C\\mid X)$ encodes that confound and whether predictions degrade when the confounder channel is removed. If the model performs the same when its 'confounder' is replaced by an arbitrary latent dimension, the causal intervention claim is not doing the work.","tokens_in":19048,"feed_emoji":"🚗","tokens_out":8920,"duration_ms":85636,"temperature":0.7,"pith_summary":"The paper tries to establish that a trajectory-prediction model can keep learning new driving and pedestrian scenes without forgetting old ones by treating each scene's bias as a latent confounder and intervening on it. It proposes C2INet, a plug-and-play module that estimates the confounder from observed trajectories, applies a backdoor-adjusted variational objective, and stores a learned prior for every scene in a bounded memory queue. The central claim is that this reduces catastrophic forgetting and improves average prediction error on ETH-UCY, a synthetic circle-crossing dataset, and the Stanford Drone Dataset, against counterfactual, invariant, and continual-learning baselines. If correct, it would let autonomous-driving and surveillance systems adapt to new regions using only current data plus a compact set of scene priors, without full retraining.","feed_headline":"Causal fix curbs forgetting in scene-by-scene trajectory prediction","feed_subtitle":"A memory queue of scene priors keeps past accuracy intact while new scenes arrive, across three datasets.","key_machinery":"The machinery is a backdoor-adjusted variational objective, meaning an objective that estimates the interventional distribution $P(Y \\mid \\mathrm{do}(X))$ by conditioning on and averaging over the confounder, with a continually updated prior. The named object is the prior queue: a memory of pseudo-features $U_i$ whose encodings $Q(C\\mid U_i)$ approximate the scenario-specific distribution of the confounder $P(C)$. Each new task adds a pseudo-feature, the mixture weights $\\alpha_k$ are optimized alternately with the prior in a min-max loop, and a pruning rule drops redundant pseudo-features once the queue exceeds capacity $\\gamma$. The causal engine is the Gaussian product that merges the trajectory encoder and confounder encoding into $P(Z\\mid X,C)$, so predictions are made from the intervened representation.","core_discovery":"On the paper's own terms, the discovery is that the confounding influence of scene-specific context $C$ on trajectory predictions can be removed continuously, not just in a fixed training set. The authors derive an evidence lower bound for the interventional distribution $P(Y \\mid \\mathrm{do}(X))$ using a variational encoder $Q(C\\mid X)$, and combine the trajectory encoder's Gaussian $P(Z\\mid X)$ with the confounder posterior into a mixed latent distribution $P(Z\\mid X,C)$ used for prediction. They then keep a representative prior for each task in a memory queue, update it in a min-max loop with mixture weights, and prune redundant pseudo-features when the queue grows too large. In their experiments, C2INet in online and offline modes often delivers the best average ADE/FDE and shows less performance decay on previously seen tasks than the compared baselines.","pith_inferences":["The paper leaves implicit that its causal story is only as strong as the identifying assumption; a safer reading is that the memory queue plus KL regularization is the reliable part, and the confounder encoder may simply be a useful auxiliary variable.","A concrete next test would be a controlled synthetic domain shift with a known ground-truth confound, checking whether $C$ recovers it and whether removing it is necessary for the reported gains.","The task-order analysis implies a practical curriculum rule: onboard noisier scenes earlier in the sequence, because later forgetting then dampens their harmful bias.","The bounded prior queue points toward a deployment pattern where each new city or agent class gets one slot in the queue, making adaptation a matter of learning and pruning a single pseudo-feature."],"forward_implications":["If the central claim is right, continual trajectory prediction can adopt causal-debiasing tools that were designed for fixed datasets, without requiring full access to old data.","The ablation results imply that replay memory alone is not enough: keeping priors diverse through symmetric KL divergence and optimizing mixture weights both contribute to the gain.","Because the module is plug-and-play, existing RNN- and CNN-based trajectory encoders can be converted to the continual setting by adding the confounder encoder and prior queue.","Storing pseudo-features rather than raw trajectories keeps memory cost tied to the number of scenes, which would matter for deployment on vehicles or edge devices."],"supporting_citations":[{"why":"Supplies the do-calculus and backdoor-adjustment rules that the intervention derivation relies on.","marker":"Pearl (2012)"},{"why":"Provides the variational reparameterization trick used to sample the confounder encoding $Q(C\\mid X)$.","marker":"Kingma (2013)"},{"why":"Motivates aggregating posteriors into a learned prior and warns that storing all data is impractical, which leads to the pseudo-feature queue.","marker":"Tomczak & Welling (2018)"},{"why":"Contributes the continual VAE boosting objective and symmetric KL regularization that the min-max prior updates build on.","marker":"Egorov et al. (2021)"},{"why":"Is the counterfactual trajectory-debiasing baseline that C2INet extends to the continual setting and compares against.","marker":"Chen et al. (2021a)"},{"why":"Is the invariant motion-forecasting baseline whose empirical-risk penalty is contrasted with the backdoor-adjustment approach.","marker":"Liu et al. (2022)"},{"why":"Supplies the generative causal representation continual-learning baseline GCRL used as a direct competitor.","marker":"Bagi et al. (2023)"},{"why":"Supplies the EWC continual-learning baseline that anchors the catastrophic-forgetting comparison.","marker":"Kirkpatrick et al. (2017)"},{"why":"Supplies the random coresets replay baseline that anchors the memory-based continual-learning comparison.","marker":"Bachem et al. (2015)"}],"fun_headline_variants":["Continual causal intervention debiases scene-wise trajectory prediction","Memory-queued priors keep trajectory accuracy across new scenes","Causal debiasing curbs forgetting in incremental trajectory models","Scene-aware causal priors make trajectory prediction generalizable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single hidden 'environment' variable, learned only from the observed trajectories, really captures every scene-specific bias, and that the causal-adjustment formulas used to remove it are valid for this model. If that premise is wrong, the intervention step is just an ordinary variational encoder with a stored prior, and the gains may come from the replay memory and regularization instead.","fun_headline_variants_meta":{"raw":{"variants":["Continual causal intervention debiases scene-wise trajectory prediction","Memory-queued priors keep trajectory accuracy across new scenes","Causal debiasing curbs forgetting in incremental trajectory models","Scene-aware causal priors make trajectory prediction generalizable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000472,"raw_usage":{"total_tokens":2341,"prompt_tokens":936,"completion_tokens":1405,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":1337}},"tokens_in":552,"tokens_out":1405,"duration_ms":10205,"temperature":1.0,"reasoning_tokens":1337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:40:42.886409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model on a synthetic dataset where the ground-truth confound is known, such as a rule that shifts all right turns in one scene, then test whether the learned $Q(C\\mid X)$ encodes that confound and whether predictions degrade when the confounder channel is removed. If the model performs the same when its 'confounder' is replaced by an arbitrary latent dimension, the causal intervention claim is not doing the work.","supporting_citations":[],"review_version":1}