{"id":"e6d5375d-3141-489f-9753-e6154bf348a2","arxiv_id":"2501.07711","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A GAN predictor that multiplies graph attention by random edge weights reports strong ETH/UCY results, but the random-weight mechanism is neither specified nor ablated.","lead":"DTGAN is a GAN-based pedestrian trajectory predictor that multiplies graph attention coefficients by random edge weights and reports lower ADE/FDE than several baselines on ETH/UCY. A smart generalist should read it because it tests whether replacing hand-made interaction rules with random weights plus a Gaussian task loss improves trajectory forecasting, but the paper does not isolate the effect of the random weights.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random-weight mechanism is undefined: the paper never states whether e_ij are fixed, resampled, or trained; if fixed they cannot 'learn' interactions, if trained they are not random, so the central claim is untestable.","rationale":"The reader's REJECT verdict remains appropriate, and the strongest shared concern is that the random-weight mechanism is underspecified. I cannot fully agree with the reader's statement that the random multipliers were never removed, because Figure 5 contains an 'All Ones' condition that is close to a standard GAT ablation; the deeper problem is that the paper does not say whether e_ij are fixed, resampled, or trained, and the text contradicts itself (Section III-A vs. Section II). If e_ij are fixed noise, the claim that they 'learn' or 'align with training' is false; if they are trained, they are not random at inference and the novelty is mislabeled. The proposed test settles this by comparing fixed-random, trained-random, and all-ones variants while logging gradients to e_ij. A secondary reporting issue: the abstract's 16.7% and 39.3% improvements are computed as (0.42−0.36)/0.36 and (0.85−0.61)/0.61, i.e., relative to the improved value rather than the baseline, which inflates the stated gains. This does not change the verdict but reinforces the need for corrected statistics. The paper's numerical system may work, but as written the central contribution is not testable, so the reader's REJECT is justified.","tokens_in":19770,"tokens_out":10591,"duration_ms":108691,"concrete_test":"Implement the three variants of Eq. (3) under the exact Section IV-B protocol: (a) e_ij fixed at independently sampled U(0,1) values (the literal reading of Section III-A); (b) e_ij trainable parameters initialized with the same U(0,1) sample; (c) e_ij = 1 for all off-diagonal edges (All Ones). Run each across the five ETH/UCY splits with 10 random seeds, record mean±std ADE/FDE, and log whether ∂L/∂e_ij is nonzero in (b). If (a) is not significantly better than (c), the claim that fixed random weights encode social interactions is false; if only (b) beats (c), the 'random' part is just initialization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that random edge weights e_ij in Eq. (3) automatically capture implicit social interactions. Section III-A states 'At is a randomly sampled weights matrix' and defines e_ij = z for z a random value, but no distribution, range, or sampling schedule is given. Section II then says 'These weights are not fixed but are aligned with the model's training, allowing the representation of social interactions to evolve dynamically with training epochs.' These descriptions are mutually incompatible. If e_ij are fixed random constants, then Eq. (4) multiplies learned attention coefficients by static noise; gradients do not flow into e_ij, so the weights cannot 'align with training' or encode learned social information. If e_ij are trainable parameters, then they are not random after initialization, and the claim that DTGAN 'eliminates pre-defined rules' reduces to standard learned edge weights with a random initializer. Figure 5 does include an 'All Ones' condition, which approximates a standard GAT baseline, so the random multipliers were not entirely unremoved; however, that comparison is a single run with no reported variance or significance test, and the definition of 'All Ones' is ambiguous regarding the diagonal. Because the core mechanism can be interpreted in two incompatible ways, the paper's central contribution cannot be verified or reproduced as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DTGAN, a GAN-based framework for pedestrian trajectory prediction in which random edge weights are introduced into graph attention. The generator processes graph sequences through spatial embedding, GAT, TCN, CNN, and a decoder, while the discriminator uses an SPE-LSTM-FC architecture. The paper explores three task-loss variants (MSE, Gaussian negative log-likelihood, and uniform negative log-likelihood), reports ADE/FDE and AMD/AMV results on ETH/UCY, and claims that DTGAN-G achieves 0.36/0.61 average ADE/FDE, outperforming all listed baselines because the random-weight mechanism automatically captures implicit social interactions without pre-defined rules.","tokens_in":20073,"tokens_out":9435,"duration_ms":88854,"significance":"If the mechanism and the quantitative claims were reliable, the paper would offer a simple modification to GAT-based trajectory predictors and a useful empirical study of task losses in adversarial training. The manuscript has strengths: comparisons against a broad set of baselines, a component ablation in Table V, a random-seed robustness study in Table IV, and qualitative analyses in Figures 3 and 4. However, the central random-weight mechanism is not specified in a way that can be implemented or tested, and several headline numbers in the abstract and Section IV-D are inconsistent with Table II. The scientific contribution therefore cannot currently be assessed as stated.","major_comments":[{"comment":"The central mechanism is not defined. Section III-A states that e_{ij}=z for i≠j with z a random value and calls A_t a randomly sampled matrix, but no distribution, range, or sampling schedule is given. Section II says 'These weights are not fixed but are aligned with the model's training, allowing the representation of social interactions to evolve dynamically with training epochs', which is incompatible with fixed random values. If e_{ij} are fixed random constants, gradients do not flow into them, so Eq. (4) multiplies learned attention coefficients by static noise and the weights cannot 'learn' or evolve; if e_{ij} are trainable parameters, they are only randomly initialized and the claim that random weights eliminate pre-defined rules reduces to standard learned edge weights. The manuscript must specify the exact mechanism and experimentally distinguish frozen random, resampled, and trainable-random-initialization variants.","section":"Section III-A, Eq. (3), and Section II"},{"comment":"The comparison of weight-generation schemes does not isolate the random-weight contribution. The All Ones condition approximates standard GAT attention, but no standard GAT baseline without the multiplicative mask is reported, and no repeated runs or error bars are given for these four variants. Table IV reports random-seed variation only for the full DTGAN-G model. The gap between Random Weights (0.36/0.61) and All Ones (0.40/0.75) is therefore not shown to be statistically meaningful, so the claim that random weights specifically capture implicit social interactions is not supported.","section":"Section IV-F, Fig. 5"},{"comment":"The claimed improvements of 16.7% for ADE and 39.3% for FDE do not correspond to any pairwise comparison in Table II. From DTGAN to DTGAN-G, the average changes are 0.42->0.36 (14.3%) and 0.85->0.61 (28.2%); against Social-STGCNN the changes are 18.2% and 18.7%. The authors should identify the reference model and correct the percentages.","section":"Section IV-D and abstract"},{"comment":"The claim that DTGAN-G gives 'the best performance among all baselines' rests on small margins over TPNMS and SocialDualCVAE (ADE 0.36 vs 0.38 and 0.38; FDE 0.61 vs 0.73 and 0.69), yet the headline results are reported without standard deviations, confidence intervals, or multiple-run statistics. Since the robustness analysis in Table IV exists, it should be extended to the final reported numbers and to the comparison variants in Fig. 5.","section":"Table II and Section IV-D"},{"comment":"The ablation study varies the GAT/GCN encoder, the SPE layer, and the temporal/decoder components, but it never removes or changes the random-weight mask. As a result, the component analysis does not test whether the random-weight term is responsible for any of the reported gains, which is the paper's central claim.","section":"Section IV-H, Table V"}],"minor_comments":[{"comment":"The sentence 'social interactions occur between more than two individuals' should read 'between pairs of individuals' or 'among more than two individuals'; as written, it conflicts with the zero diagonal and the off-diagonal definition of e_{ij}.","section":"Section III-A"},{"comment":"The notation TpY appears to be a misrendered product over t=1,...,Tp; please correct the product symbol and specify the parameters over which the argmin is taken.","section":"Section III-D, Eq. (13)"},{"comment":"The condition sqrt(xt_i^2 + yt_i^2) < r_hat uses absolute coordinates; it should use displacement from the predicted center of the circular distribution. Please clarify.","section":"Section III-D, Eq. (12)"},{"comment":"The sentence about WGAN cites 'Huang et al. [52]' but reference [52] is Lucic et al., 'Are GANs created equal?'; the citation should be corrected.","section":"Section III-D"},{"comment":"Pre-training is mentioned as a separate stage, but no pre-training loss, number of epochs, or stopping criterion is provided; these details are needed for reproducibility.","section":"Section IV-B"},{"comment":"The caption says each density curve represents a distribution of values, but the figure shows line plots of ADE/FDE versus random seeds; please align the caption with what is actually plotted.","section":"Figure 6"},{"comment":"The text 'Unique, Group, Len, Index and Append are specific function functions' contains a typo, and minPed is used in the algorithm but is not defined in the notation table. Please fix both.","section":"Algorithm 1"}],"recommendation":"reject","confidential_remarks":"The manuscript header indicates an acceptance date of 12 February 2024 and a current version date of 24 April 2024, while the arXiv submission is dated 13 January 2025; the editor may wish to check prior-publication or overlap status. This did not affect my technical assessment. In my view, the core problem is the incoherence of the random-weight mechanism: as written, it is either static noise multiplied into learned attention or merely a random initialization of trainable edge weights, so the central contribution is not a well-defined, testable mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest empirical paper in a crowded subfield. It reports strong ADE/FDE numbers on ETH/UCY with a GAN-GAT hybrid, and it includes a robustness analysis over random seeds and an ablation over loss functions. The one genuinely new ingredient is multiplying learned GAT attention coefficients by a per-edge random weight (Eq. 3). That is a narrow variation, but it is a real one.\n\nWhat the paper does well: the experimental work is more thorough than the typical submission. Table IV reports mean and std over five random seeds, and the variance is small. Fig. 5 compares four weight schemes, including an all-ones matrix that approximates standard GAT, and random weights come out best. The loss exploration (WGAN with MSE, Gaussian NLL, uniform NLL) is a reasonable empirical study, and the Gaussian NLL variant is clearly better than the others. The ablation over generator/discriminator components is standard but carried out on all datasets.\n\nThe soft spots are significant. The central claim—that random weights automatically capture implicit social interactions—is not actually tested. The all-ones baseline is the closest thing to removing the random multiplier, but it is a single run, and the paper never compares against plain GAT with no edge multiplier or with learned edge weights. So the attribution of the improvement to randomness specifically is not established. Second, the paper never specifies the distribution of the random weights, whether they are fixed during training, resampled, or trained. Section III-A says At is randomly sampled; Section II says the weights 'are not fixed but are aligned with the model's training.' Those are incompatible readings, and neither is backed by a gradient update rule or a sampling schedule. Third, the abstract claims improvements of 16.7% and 39.3% without stating the baseline; the numbers in Table II and Table III do not obviously produce those percentages. Finally, there is a garbled sentence in the loss section attributing a claim about GAN variants to 'Huang et al.' when the reference is Lucic et al. These are fixable in revision.\n\nBottom line: the paper is not a breakthrough, but it is a real empirical system with useful ablations and stable reported numbers. The central mechanism is under-specified and the attribution is unsupported as written, but the issues are addressable. I would give it a serious referee rather than desk reject, because there is enough substance to warrant a careful read and the authors can fill the gaps. For my own work, I wouldn't cite it until the weight scheme is clarified.","headline":"A modest empirical contribution whose central random-weight mechanism is under-specified and not properly ablated, but with enough solid experimental work to deserve a referee rather than a desk reject.","tokens_in":20592,"tokens_out":2677,"would_cite":false,"duration_ms":25108,"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":"Multiplying graph attention by random edge weights lets a GAN learn pedestrian interactions without predefined rules, reaching 0.36/0.61 ADE/FDE on ETH/UCY.","keywords":["pedestrian trajectory prediction","social interaction modeling","graph random weights","generative adversarial network","graph attention network","Gaussian negative log-likelihood","ETH/UCY benchmark","multi-modal trajectory forecasting"],"falsifier":"Train DTGAN-G identically on ETH/UCY with every edge multiplier fixed to 1 (a plain GAT) and compare ADE/FDE with the random-weight version. If the two are statistically indistinguishable, or if the plain-GAT version matches the reported 0.36/0.61, then the random weights are not carrying the interaction information; a complementary check is that the learned attention after training concentrates on pedestrians ahead or on collision course regardless of the sampled $e_{ij}$ values, which would mean the network learns interactions from coordinates and the random matrix is uninformative noise.","tokens_in":19566,"feed_emoji":"🚶","tokens_out":10710,"duration_ms":93565,"temperature":0.7,"pith_summary":"Pedestrian trajectory prediction usually models social interaction by hand-coded rules such as Euclidean distance between walkers. The paper claims those rules are biased and incomplete: a target pedestrian may pay less attention to a nearby person behind them than to someone ahead they might collide with. To remove the rule, this work proposes DTGAN, a GAN that treats the observed scene at each time step as a graph whose edge weights are random values, multiplied into learned graph-attention coefficients before temporal decoding. With a Gaussian negative-log-likelihood task loss added to a Wasserstein adversarial objective, the best variant DTGAN-G reports average ADE/FDE of 0.36/0.61 meters on the ETH/UCY benchmark, ahead of listed baselines including Social-STGCNN at 0.44/0.75. The claim that matters: if random weights work, then interaction structure does not need to be specified in advance, and a network can discover it from data alone.","feed_headline":"Random-weight graph model cuts pedestrian-track error to 0.36 m","feed_subtitle":"Replacing hand-coded interaction rules with random edge weights plus a Gaussian loss beats listed baselines on ETH/UCY.","key_machinery":"The load-bearing mechanism is the random-weight attention multiplication of Eq. (3), $\\hat\\alpha^t_{ij} = \\alpha^t_{ij} \\cdot e^t_{ij}$, where $\\alpha^t_{ij}$ is the standard GAT attention coefficient from Eq. (2) and $e^t_{ij}$ is a scalar edge weight attached to the pair of pedestrians $(i,j)$ at time $t$. The paper calls the matrix of these values a randomly sampled adjacency with zeros on the diagonal, and the related-work discussion adds that the weights are not fixed but evolve with training. That multiplication is the piece claimed to free the model from pre-defined interaction rules: none of the $e^t_{ij}$ values is a function of distance or direction, so the network cannot lean on a static rule and must learn which neighbors matter. The rest of the pipeline, including the temporal convolutional network, convolutional decoder, Wasserstein adversarial objective, and the chosen task loss, carries these randomly weighted node features into multi-modal trajectory predictions.","core_discovery":"The central claim, stated on the paper's own terms, is that a randomly weighted adjacency matrix inside a graph attention network is enough to capture implicit social interactions among pedestrians. At each observed time step the model builds a graph $G_t = (V^t, A^t)$ whose nodes are pedestrians and whose adjacency entries satisfy $e^t_{ij} = z$ for $i \\neq j$, where $z$ is a random value and the diagonal is zero. In Eq. (3) the learned GAT attention coefficient $\\alpha^t_{ij}$ is multiplied by $e^t_{ij}$ to give $\\hat\\alpha^t_{ij}$, and the resulting node features pass through a temporal CNN, a convolutional decoder, and adversarial training. The paper contrasts this with Social-STGCNN's fixed reciprocal-distance weights, arguing that random weights remove the bias of rules like 'nearer means more important' and let the network assign importance adaptively. Tested on ETH and UCY with 8 observed frames and 12 predicted frames, the Gaussian negative-log-likelihood variant DTGAN-G reports average ADE/FDE of 0.36/0.61, which the authors read as evidence that the model understands pedestrian intent, especially in crowded UCY scenes.","pith_inferences":["The random multipliers may be doing something more mundane than 'capturing implicit interactions': as fixed random noise they would act as a stochastic mask that regularizes attention, and as trainable parameters they would just be learned edge weights with random initialization; the paper's experiments do not distinguish these roles.","A direct way to test the paper's own story is to keep DTGAN-G but replace $e_{ij}$ with a learned function of relative position and velocity; if that model matches or beats random weights, then a data-driven rule, not randomness, is the active ingredient.","The same recipe, multiplying graph attention by random edge weights, could transfer to other GNN tasks that currently use predefined adjacency, such as multi-agent motion or traffic forecasting, but evidence would be needed that the benefit is not specific to the ETH/UCY setup."],"forward_implications":["If the random-weight mechanism works as claimed, graph-based interaction modeling in trajectory prediction no longer needs a hand-coded weighting function such as reciprocal distance, removing a source of scene-specific bias.","Adding a distributional task loss (Gaussian negative log-likelihood) to the adversarial objective yields the paper's largest measured gains: 16.7% on ADE and 39.3% on FDE relative to the plain DTGAN variant.","Because DTGAN treats each time step as a graph and decodes with TCN/CNN, adversarial training can be extended from flat sequences to graph-sequence data, which the paper frames as a new use of GANs.","On crowded UCY scenes with more social interactions, the reported gains over GAN-based baselines in ADE/FDE are larger than on ETH, which the paper reads as evidence that graph-structured node relationships help in complex crowds.","The robustness experiment across five random seeds reports small standard deviations, so the method's performance does not appear to depend on one lucky draw of the random matrix."],"supporting_citations":[{"why":"Supplies the graph-based social interaction baseline whose hand-coded reciprocal-distance weights motivate the random-weight design and are compared against in Table II and Fig. 5.","marker":"[23]"},{"why":"Defines the adversarial multi-modal trajectory prediction setup, including best-of-K sample evaluation, that DTGAN extends to graph sequences.","marker":"[19]"},{"why":"Provides the Wasserstein GAN objective used as DTGAN's adversarial loss in Eq. (9).","marker":"[51]"},{"why":"Defines the graph attention network that produces the attention coefficients multiplied by random weights in Eq. (3).","marker":"[45]"},{"why":"Establishes the social-pooling baseline, the leave-one-out evaluation protocol, and the ADE/FDE metrics used across the paper.","marker":"[16]"},{"why":"Supplies the ETH dataset used for training and evaluation in the experiments.","marker":"[56]"},{"why":"Supplies the UCY dataset used for training and evaluation in the experiments.","marker":"[57]"}],"fun_headline_variants":["Random edge weights learn pedestrian intent, beat fixed rules","GAN with random weights predicts paths, ADE 0.36 m","Random-weight graph cuts pedestrian prediction error to 0.36 m","No predefined rules: random weights capture social cues in crowds","Implicit social interactions via random weights: 0.36 m ADE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the random edge multipliers in Eq. (3), not the architecture, the adversarial loss, or the task loss, are what let the model capture implicit social interactions; the paper never isolates this premise because its comparisons swap random weights for Euclidean, all-ones, or arithmetic-progression matrices without ever running a plain GAT, and it never states the distribution of $z$ or whether $z$ is updated during training.","fun_headline_variants_meta":{"raw":{"variants":["Random edge weights learn pedestrian intent, beat fixed rules","GAN with random weights predicts paths, ADE 0.36 m","Random-weight graph cuts pedestrian prediction error to 0.36 m","No predefined rules: random weights capture social cues in crowds","Implicit social interactions via random weights: 0.36 m ADE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1563,"prompt_tokens":1003,"completion_tokens":560,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":471}},"tokens_in":619,"tokens_out":560,"duration_ms":5396,"temperature":1.0,"reasoning_tokens":471,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:37:13.050841+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DTGAN-G identically on ETH/UCY with every edge multiplier fixed to 1 (a plain GAT) and compare ADE/FDE with the random-weight version. If the two are statistically indistinguishable, or if the plain-GAT version matches the reported 0.36/0.61, then the random weights are not carrying the interaction information; a complementary check is that the learned attention after training concentrates on pedestrians ahead or on collision course regardless of the sampled $e_{ij}$ values, which would mean the network learns interactions from coordinates and the random matrix is uninformative noise.","supporting_citations":[{"cited_title":"Social-stgcnn: A social spatio-temporal graph convolutional neural network for human trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"Supplies the graph-based social interaction baseline whose hand-coded reciprocal-distance weights motivate the random-weight design and are compared against in Table II and Fig. 5."},{"cited_title":"Social gan: Socially acceptable trajectories with generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Defines the adversarial multi-modal trajectory prediction setup, including best-of-K sample evaluation, that DTGAN extends to graph sequences."},{"cited_title":"Wasserstein generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Provides the Wasserstein GAN objective used as DTGAN's adversarial loss in Eq. (9)."},{"cited_title":"Graph attention networks,","cited_arxiv_id":null,"evidence_quote":"Defines the graph attention network that produces the attention coefficients multiplied by random weights in Eq. (3)."},{"cited_title":"Social lstm: Human trajectory prediction in crowded spaces,","cited_arxiv_id":null,"evidence_quote":"Establishes the social-pooling baseline, the leave-one-out evaluation protocol, and the ADE/FDE metrics used across the paper."},{"cited_title":"You’ll never walk alone: Modeling social behavior for multi-target tracking,","cited_arxiv_id":null,"evidence_quote":"Supplies the ETH dataset used for training and evaluation in the experiments."},{"cited_title":"Learning an image-based motion context for multiple people tracking,","cited_arxiv_id":null,"evidence_quote":"Supplies the UCY dataset used for training and evaluation in the experiments."}],"review_version":1}