{"id":"1050927f-22a0-4d28-b49d-a05518247760","arxiv_id":"2501.14266","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TrajFlow predicts each future position of an agent as a separate probability cloud using normalizing flows and neural differential equations, with mixed benchmark results.","lead":"A new model called TrajFlow predicts where moving cars and people will be at each future moment by drawing a probability cloud of possible locations, instead of predicting one fixed path. Why it matters: the model can output both occupancy maps and trajectory samples in continuous time, which are the two representations self-driving cars and traffic systems use for planning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The marginal-vs-joint accuracy claim is evaluated with minADE/minFDE on independently sampled per-timestep locations (Eq. 11), a protocol that rewards incoherent trajectories, and Table 1 itself shows the joint model is more accurate on ETH/UCY.","rationale":"I read the paper in good faith and identify the same load-bearing weakness the reader flagged. The framework is a coherent combination of known components, and the marginal occupancy density modeling is a legitimate contribution. However, the headline claim that the marginal formulation produces higher trajectory-forecasting accuracy is undermined by a mismatch between the generative model and the evaluation metric. The model trained with Eq. 11 does not produce joint trajectories; it produces independent per-timestep locations. minADE/minFDE are trajectory-level metrics designed for coherent multi-modal trajectory samples, and taking the minimum over 20 independent per-step samples systematically favors the marginal model by allowing it to avoid committing to one mode or path. This is more than a philosophical objection: Table 1 shows the joint version of the same architecture beats the marginal version on both minADE and minFDE on ETH/UCY. The inD experiment avoids the issue by using RMSE and CRPS, which are per-point metrics, but then those results do not support the trajectory-forecasting claim. The paper's own discussion concedes the diversity limitation, and the top-k heuristic is a post-hoc patch, not a principled joint model. The concern is concrete and testable: compare full-trajectory likelihoods. If the marginal model has worse trajectory-level likelihood but better minADE, the reported accuracy advantage is an artifact. I do not think this requires rejection, because the marginal occupancy estimation results and the architectural ablations remain informative, and the trajectory-forecasting claim can be revised or re-evaluated. The reader's conditional verdict is appropriate, so I leave it unchanged.","tokens_in":15410,"tokens_out":6106,"duration_ms":62682,"concrete_test":"Run a paired evaluation on the same ETH/UCY leave-one-out folds: for CDE-CNF (Marginal) and CDE-CNF (Joint), compute the full-trajectory negative log-likelihood on held-out trajectories, using the product density prod_s p(u_s|O) for the marginal model and the joint flow density for the joint model. If the marginal model's trajectory NLL is higher (worse) than the joint model's while its minADE is lower, that demonstrates the minADE advantage is an artifact of independent per-step sampling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the marginal formulation yields higher trajectory-forecasting accuracy is not supported by the paper's own evidence and rests on an evaluation mismatch. Training uses Eq. 11, P(U|O) = prod_s p(u_s|O), so the model generates a trajectory by sampling each future position independently from its marginal density. The paper then scores those samples with trajectory-level minADE/minFDE in the Experiments section. Because the minimum is taken over 20 samples drawn independently per timestep, the marginal model can exploit combinatorial diversity: the selected sample can pick the most probable location at each time without satisfying any temporal coupling. This artificially lowers minADE/minFDE relative to joint models that must commit to a coherent path. The internal comparison in Table 1 makes the problem visible: CDE-CNF (Joint) achieves 0.18/0.37 while CDE-CNF (Marginal) achieves 0.19/0.38 on ETH/UCY, contradicting the abstract's statement that the marginal formulation gives higher accuracy. The inD RMSE/CRPS results in Table 2 evaluate per-point marginal accuracy rather than trajectory quality, so they do not validate the trajectory-forecasting claim either. Thus the state-of-the-art trajectory-forecasting claim is not established; the paper's valid contribution is marginal occupancy density estimation, not superior trajectory prediction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TrajFlow, a generative framework for occupancy density estimation. The model uses a causal encoder (GRU or neural CDE) to embed an observed trajectory and a normalizing flow (affine couplings or neural ODE) to model the conditional marginal density of each future spatial location p(u_s | O, F). The authors claim that modeling marginal rather than joint trajectory densities yields higher forecasting accuracy, enables continuous-time sampling, and supports both trajectory sampling and occupancy grids. They evaluate discrete and continuous implementations on ETH/UCY and inD, reporting ablations and comparisons against six baselines.","tokens_in":15631,"tokens_out":4283,"duration_ms":36023,"significance":"If the claims were fully supported, the paper would provide a useful contribution to motion forecasting by unifying trajectory and occupancy representations in a single flow-based model. The code is publicly available, and the systematic ablation of discrete versus continuous encoders and flows is informative. The core methodological idea of marginal occupancy density estimation is worthwhile. However, the paper's central claim that the marginal formulation produces higher trajectory-forecasting accuracy is not supported by its own evidence, and the evaluation protocol for trajectory metrics is mismatched with the marginal sampling procedure. The contribution is better positioned as occupancy-density estimation than as superior trajectory prediction.","major_comments":[{"comment":"The abstract and contribution list claim that the marginal formulation produces higher accuracy on trajectory forecasting benchmarks, but Table 1 shows CDE-CNF (Marginal) at 0.19/0.38 versus CDE-CNF (Joint) at 0.18/0.37 on ETH/UCY, and the Discussion itself states that 'the marginal and joint formulations produced comparable results on the ETH/UCY dataset.' This internal contradiction must be reconciled; either the claims should be revised or additional evidence should be provided to support the stated advantage.","section":"Abstract; Table 1; Discussion"},{"comment":"Because the model generates trajectories by independently sampling each future position from the marginal density (Eq. 11), the minADE/minFDE metrics, which select the best of 20 independently drawn samples, can reward temporally incoherent trajectories: the best sample can pick the most probable location at each time step without forming a physically plausible path. The inD results (Table 2) use RMSE/CRPS, which are per-point marginal metrics and do not assess trajectory coherence. To support the trajectory-forecasting claim, the authors should evaluate with trajectory-level metrics that respect temporal coupling, for example by sampling with a coherence-enforcing strategy (such as their top-k sampling) or by measuring the error of full sampled trajectories without per-timestep recombination.","section":"Eq. (11) and Evaluation Metrics"},{"comment":"The inD evaluation uses a single random 75/25 split of one recording session, which is nonstandard and does not permit any assessment of variance; this makes the reported state-of-the-art claim fragile. The authors should adopt the standard inD benchmark splits or report results across multiple seeds/splits with confidence intervals, and should clarify how the reported baseline numbers were obtained under the same protocol.","section":"Vehicle Experiment"},{"comment":"The Discussion concedes that the marginal formulation 'inherently reduces diversity among the sampled motion trajectories and introduces noise, as the relationships between predicted locations are no longer explicitly modeled,' and Figure 6 shows that a top-k sampling strategy is needed to reduce the resulting noise. This admission directly undermines the abstract's claim that the marginal formulation produces higher trajectory-forecasting accuracy, and it should be reflected in the paper's framing: the contribution is better positioned as occupancy-density estimation, not as superior trajectory prediction.","section":"Discussion and Figure 6"}],"minor_comments":[{"comment":"The text calls the dataset 'UTY' but it should be 'UCY'; please correct this typo.","section":"Experiments (pedestrian dataset description)"},{"comment":"In the paragraph discussing inD baseline comparison, 'MSE' should be 'RMSE' to match the metric defined in the Evaluation Metrics subsection.","section":"Baseline Comparison (inD)"},{"comment":"The paper does not provide sufficient architecture details for reproducibility, such as the number of affine coupling layers, hidden sizes and activation functions for the MLPs in the coupling layers and the CDE vector field, the FiLM layer dimensions, and training hyperparameters like batch size; please add these details.","section":"Methodology / implementation details"},{"comment":"Equation (32) is not typeset correctly; the max over u_i,s is missing its intended subscript and the expression should be defined more carefully so the additive fusion procedure is unambiguous.","section":"Discussion, Eq. (32)"},{"comment":"The text calls Eq. (27) the 'concat and squash layer', but the equation describes FiLM conditioning; the terminology is confusing and should be aligned with the FiLM citation.","section":"Methodology, Eq. (27)"},{"comment":"The claim of state-of-the-art performance should be qualified: in Table 1, CDE-CNF does not beat PPT on minFDE (0.38 vs. 0.31), so the state-of-the-art statement is too strong as written.","section":"Contributions, item 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not circular; the comparisons are out-of-sample and the code release is a strength. The main issue is overclaiming: the marginal formulation's advantage is presented as accuracy on trajectory forecasting, which the supplied evidence refutes. With a revised framing centered on occupancy density estimation and a corrected evaluation that either uses coherent trajectory sampling or explicitly limits the claims to marginal density quality, the paper could be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read TrajFlow alongside your notes. I think the stress-test is right: the paper's headline claim about trajectory accuracy doesn't survive its own results. Table 1 has CDE-CNF (Joint) at 0.18/0.37 and CDE-CNF (Marginal) at 0.19/0.38 on ETH/UCY, so the marginal formulation is worse on the trajectory metrics there. The abstract's 'higher accuracy' and the contribution bullet claiming state-of-the-art performance are overstated. And your point about Eq. 11 is the mechanism: sampling each future position independently and then scoring with minADE/minFDE over 20 draws lets the marginal model pick the best location at every time step without committing to a coherent path. That is a real evaluation mismatch.\n\nWhat is actually new: modeling the per-time marginal spatial density with a normalizing flow conditioned on a neural-CDE embedding, with continuous-time sampling via a CNF. I don't see that exact combination in FloMo, Trajectron++, or SocialVAE. The paper also ships code, runs clean ablations across four configurations, and is honest in the Discussion about the joint/marginal trade-off and the diversity limitation. The inD results (Table 2) show the marginal CDE-CNF beating the joint on RMSE/CRPS, but those are per-point marginal metrics, so they support the occupancy-density claim, not the trajectory-forecasting claim. Citation pattern is clean: the only self-citation is the second author's tutorial [34], used as background. No circularity in the empirical comparisons.\n\nSoft spots in order of importance. First, the evaluation protocol: inD uses a single recording session split with no error bars, and only six baselines, so 'state-of-the-art' is not established. Second, the conditional-independence factorization is a strong assumption for trajectory generation; the paper acknowledges reduced diversity, but the evaluation still treats the samples as trajectories. Third, the runtime costs are severe (CDE-CNF is 19x slower to train than GRU-DNF), which undercuts the practical contribution unless the occupancy-grid use case is the target. None of these sink the core idea; they mean the claims need to be revised.\n\nWho this is for: researchers working on occupancy grids and continuous-time trajectory density models, especially in motion forecasting. The paper deserves a serious referee: the construction is coherent, the code is public, and the marginal-vs-joint question is worth airing. My recommendation: send it to review, but the authors should be asked to fix the abstract, re-run or re-report the trajectory comparison with a standard inD protocol, and add a discussion of the sampling-diversity/evaluation mismatch. If they can't soften the claims, the paper is still a decent workshop-level contribution, but as written the central accuracy claim is not supportable.","headline":"TrajFlow's marginal occupancy density model is a real and useful construction, but the paper's SOTA trajectory-forecasting claim is contradicted by its own Table 1 and an evaluation protocol that rewards incoherent samples.","tokens_in":16241,"tokens_out":3141,"would_cite":true,"duration_ms":25326,"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":"TrajFlow claims that predicting each future location's marginal density with a normalizing flow, instead of the joint density of the full trajectory, yields state-of-the-art forecast accuracy and enables fully continuous occupancy sampling.","keywords":["trajectory forecasting","occupancy density estimation","normalizing flows","neural ordinary differential equations","neural controlled differential equations","marginal density","continuous normalizing flow","motion forecasting"],"falsifier":"Take a held-out set of real trajectories and compare the likelihood a trained marginal model assigns to each true trajectory under the product formula against the likelihood assigned by a joint flow: if the product model systematically underrates coherent trajectories or overrates zig-zag paths that hop between plausible lanes, the independence assumption is the weak point and the reported minADE gains are an artifact of how the evaluation samples trajectories.","tokens_in":15127,"feed_emoji":"🚗","tokens_out":10164,"duration_ms":82002,"temperature":0.7,"pith_summary":"TrajFlow is a generative framework for motion forecasting built on one design choice: model each future location's probability distribution separately rather than the joint distribution of the entire future path. The paper argues this marginal formulation is more accurate on standard benchmarks, allows a single model to sample locations at arbitrary future times, and directly produces the two most common motion-forecasting outputs—per-agent trajectories and occupancy grids. The authors implement the idea with neural differential equations in both components—a controlled differential equation as the causal encoder and a continuous normalizing flow as the decoder—and report that this fully continuous version outperforms six baseline models on the ETH/UCY pedestrian benchmarks and on the inD vehicle-intersection dataset. The main cost is compute: training and inference are much slower than a discrete implementation. If the claim is right, one causally encoded flow can serve both trajectory sampling and occupancy mapping for autonomous driving and intersection monitoring.","feed_headline":"Marginal normalizing flow tops trajectory forecasting benchmarks","feed_subtitle":"Predicting where an agent will be, not the whole path, gives continuous-time occupancy maps.","key_machinery":"The central object is the marginal occupancy density $p(u_s^i \\mid O_i,F_i)$, implemented as a conditional normalizing flow. The transform $u_s^i = \\Psi_{\\psi}^{-1}(\\zeta,s,z_s)$ maps a base-latent sample through an invertible network conditioned on the causal embedding and the forecast time $s$; training maximizes exact likelihood through the change-of-variables formula, with a cheap log determinant for affine coupling layers and an integrated trace for the continuous normalizing flow via the instantaneous change of variables. The causal encoder is either a GRU or a neural controlled differential equation driven by a natural cubic spline through the observed trajectory. The load-bearing identity is the factorization in Eq. 11, which makes continuous-time sampling and additive-fusion occupancy grids tractable but also removes explicit modeling of correlations between future positions.","core_discovery":"The paper's central claim is that the conditional marginal density $p(u_s^i \\mid O_i,F_i)$—the probability that agent $i$ occupies location $u_s^i$ at future time $s$ given the observed trajectory $O_i$ and derived features $F_i$—can be learned directly and is a better forecasting target than the joint density of the full unobserved trajectory. The framework factorizes the likelihood of an unobserved path as the product of per-step marginal densities, $P(U \\mid O)=\\prod_{s=1}^{S} p(u_s \\mid O)$, and trains a normalizing flow $\\Psi_{\\psi}$ that maps a latent sample $z_s$ and time $s$, conditioned on a causal embedding $\\zeta = \\Phi_{\\varphi}(O_i,F_i)$, to a location $u_s^i$. The fully continuous configuration—a neural controlled differential equation encoder and a continuous normalizing flow decoder—achieves the best reported minADE on ETH/UCY and the best RMSE and CRPS on inD among the compared baselines, and the marginal formulation outperforms the paper's own joint formulation on the long-horizon inD data while staying comparable on ETH/UCY.","pith_inferences":["Editorial inference: because Eq. 11 drops correlations between future positions, independently sampled time steps produce noisy, incoherent paths—the paper's own top-k sampling is a patch for this. A natural extension is to add a lightweight second-stage temporal model (a copula or small autoregressive head) on top of the marginal flow to restore coherence while keeping the occupancy-density benef","Editorial inference: the marginal formulation's advantage is not uniform—on ETH/UCY the joint version actually has slightly better minADE and minFDE, while on inD the marginal version is clearly better. This suggests the benefit grows with forecast horizon and scene complexity, so the strongest test of the paper's idea would be other long-horizon vehicle datasets.","Editorial inference: a calibrated marginal density can be used directly for collision-probability estimates and black-spot identification, as the paper suggests but does not test; a concrete follow-up would compare TrajFlow-derived occupancy maps against existing black-spot methods on road-safety data.","Editorial inference: the ETH/UCY results sit close to what the paper calls the entropy floor of the dataset, so the practical improvement over baselines on pedestrian scenes is small; the cleaner evidence for the marginal design is the comparison against the paper's own joint baseline under the same architecture."],"forward_implications":["A single trained TrajFlow model can output both trajectory samples and occupancy grids, because the same marginal density can be sampled per time step or fused over time with the additive-fusion rule in Eq. 32.","Occupancy grids can be generated at any sampling frequency, not just the training time step, because forecast time is an explicit input to the flow; the paper shows grids sampled at 10 times the training frequency.","The fully continuous CDE-CNF configuration reaches the reported results with far fewer parameters than the compared baselines: about 60,000 on ETH/UCY and 2.7 million on inD.","On the long-horizon inD experiment, the marginal formulation beats the paper's own joint formulation (RMSE 1.14 vs 1.47; CRPS 0.38 vs 0.45), supporting the claim that marginal densities help long-horizon forecasting.","The computational cost is substantial: the fully continuous model is roughly 19 times slower to train and 11 times slower at inference than the fully discrete GRU-DNF on the pedestrian experiment, so it suits batch or memory-limited settings rather than real-time onboard use."],"supporting_citations":[{"why":"Introduces the normalizing-flow density-estimation framework that TrajFlow's exact-likelihood training relies on.","marker":"(9)"},{"why":"Defines neural ordinary differential equations and the instantaneous change-of-variables formula used by the continuous normalizing flow.","marker":"(18)"},{"why":"Defines neural controlled differential equations, the continuous encoder used in the fully continuous configuration.","marker":"(19)"},{"why":"Provides the continuous-normalizing-flow training method and trace estimator used in the continuous decoder.","marker":"(41)"},{"why":"Supplies the affine coupling layers whose triangular Jacobian makes the discrete flow's log-determinant computation efficient.","marker":"(43)"},{"why":"Trajectron++ is the strongest generative baseline and the source of the evaluation protocol; the marginal formulation is positioned against its joint-density approach.","marker":"(7)"},{"why":"FloMo is the closest prior model, an RNN-conditioned normalizing flow for joint trajectory density, making it the direct point of comparison for the marginal formulation.","marker":"(8)"},{"why":"ETH is one of the two pedestrian trajectory datasets used in the leave-one-out evaluation.","marker":"(48)"},{"why":"UCY provides the remaining pedestrian scenes in the ETH/UCY benchmark.","marker":"(49)"},{"why":"inD supplies the vehicle-intersection dataset where the marginal formulation shows its largest advantage over the joint formulation.","marker":"(50)"}],"fun_headline_variants":["Marginal normalizing flow beats joint on trajectory benchmarks","Continuous-time occupancy maps via neural ODE flow","Predict where, not the path: flow-based occupancy density","Neural ODE flow for continuous future location sampling","Marginal flow tops joint on long-horizon forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that, once the observed past is known, each future location is independent of every other future location, and that assumption is what makes the per-step product formula tractable.","fun_headline_variants_meta":{"raw":{"variants":["Marginal normalizing flow beats joint on trajectory benchmarks","Continuous-time occupancy maps via neural ODE flow","Predict where, not the path: flow-based occupancy density","Neural ODE flow for continuous future location sampling","Marginal flow tops joint on long-horizon forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000792,"raw_usage":{"total_tokens":3535,"prompt_tokens":1037,"completion_tokens":2498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":2421}},"tokens_in":653,"tokens_out":2498,"duration_ms":17979,"temperature":1.0,"reasoning_tokens":2421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:14:57.708911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of real trajectories and compare the likelihood a trained marginal model assigns to each true trajectory under the product formula against the likelihood assigned by a joint flow: if the product model systematically underrates coherent trajectories or overrates zig-zag paths that hop between plausible lanes, the independence assumption is the weak point and the reported minADE gains are an artifact of how the evaluation samples trajectories.","supporting_citations":[],"review_version":1}