{"id":"f98f3f88-1b88-4154-9a42-1c55237002de","arxiv_id":"2504.15616","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SocialMOIF combines multi-order attention fusion, a VAE-style trajectory distribution approximator, and a KAN-based global optimizer to achieve reported state-of-the-art ADE/FDE on ETH/UCY, NBA, SDD, and NuScenes.","lead":"A new pedestrian trajectory prediction model, SocialMOIF, fuses first-order and higher-order neighbor intentions and reports lower prediction errors than prior state-of-the-art methods across four datasets. The work is a potential step for autonomous driving and surveillance systems that need accurate short-term motion forecasts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 17's direction-loss sign is internally inconsistent with Eq. 18 as written; the reported gain from the direction component cannot be reproduced without a sign correction or explicit optimization-direction statement.","rationale":"The reader's weakest_assumption concerns the decomposition of higher-order intentions into pairwise first-order interactions, which is a conceptual gap. I agree that the higher-order layer is implemented as standard multi-head self-attention over neighbors and that the paper gives no derivation or ablation isolating a genuinely higher-order mechanism. However, that gap concerns the naming and novelty of the module, not directly the validity of the empirical performance reported in Table 1. The sign inconsistency in Eqs. 16-18 is more load-bearing because it sits at the exact point where the paper claims a new loss function contributes to the SOTA numbers: if the loss is written incorrectly, the reported improvement from component A cannot be reproduced from the manuscript alone, and the central claim becomes contingent on an undocumented fix in the code. The reader did flag the sign issue in the rationale, but not as the weakest assumption, so my agreement is partial. I would keep the verdict conditional: the paper should either correct Eq. 17/18 or make the optimization direction explicit, release the code with a commit hash, and confirm the ablations with multiple seeds and variance reporting. The sign check is a single, decisive experiment that settles whether this concern is a typo or a substantive flaw.","tokens_in":13479,"tokens_out":5749,"duration_ms":58382,"concrete_test":"Fetch the released code from https://github.com/XiaodZhao/SocialMOIF and locate the implementation of the direction loss corresponding to Eq. 17. Run the same ETH/UCY training protocol with the loss exactly as written and with the sign of the arccos term flipped, keeping all other hyperparameters fixed. If the Table 1 numbers are reproduced only with the flipped sign, Eq. 17 is a typo and the empirical claim can stand after correction; if they are reproduced with the negative sign, the paper must explain how minimizing a negative angle loss improves directional alignment. A smaller scale check: compute the gradient of Eq. 18 on a single synthetic pair (predicted direction opposite ground truth) and verify which sign decreases the angle.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that the distance-direction loss contributes to the SOTA results (Table 2, groups 6 vs 7). As written, Eq. 17 defines Langle = -arccos( (\\hat g_i^t - \\hat g_i^{t+1}) · (g_i^t - g_i^{t+1}) / (norms) ). Because arccos of a cosine similarity is an angle in [0, π], the term is negative or zero. If Eq. 18 is a loss to be minimized, minimizing Langle maximizes the angle between the predicted and ground-truth direction, actively repelling the prediction from the target direction. If instead Eq. 18 is meant as an ELBO objective to be maximized, then the distance term Ldis in Eq. 16 has the wrong sign, since maximizing Ldis would increase prediction error. Thus Eqs. 16-18 do not define a coherent optimization objective under either reading. This is not a cosmetic typo: the ablation attributes a large drop (SDD from 0.20/0.27 to 0.17/0.24; NuScenes from 0.93/1.61 to 0.92/1.56) to adding the direction term, so the paper's own component analysis depends on which sign is actually used in training.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SocialMOIF, a generative trajectory prediction model whose stated contributions are a multi-order intention fusion module (first-order target-neighbor interactions plus higher-order neighbor-neighbor interactions), a trajectory distribution approximator based on latent variables, a KAN-based global trajectory optimizer for parallel decoding, and a distance-direction fused loss. The model is evaluated on ETH/UCY, NBA, SDD, and NuScenes with best-of-20 ADE/FDE metrics, and the paper reports state-of-the-art results on essentially every benchmark, with ablations attributing gains to each component. The central claim is that SocialMOIF outperforms previous state-of-the-art baselines across multiple metrics in both dynamic and static datasets.","tokens_in":13723,"tokens_out":5806,"duration_ms":54219,"significance":"If the reported numbers are reproducible, this would be a strong empirical result: the improvements on NBA (e.g., Rebound ADE 0.34 vs 0.54), SDD (ADE 0.17 vs 0.21), and NuScenes (ADE 0.92 vs 1.04) are substantial. The paper also ships a code link, compares with a wide set of baselines, and provides component-level ablations, which are strengths. However, the significance is currently conditional on resolving serious internal inconsistencies in the definition of the loss and the fusion formula, and on providing statistical support for the headline claims.","major_comments":[{"comment":"The optimization objective is internally inconsistent as written. Langle in Eq. (17) is defined as minus the arccosine of a cosine similarity, so Langle is non-positive; if L in Eq. (18) is minimized, the direction term repels the predicted displacement from the ground-truth direction, and if L is instead maximized as an ELBO, the positive distance term Ldis would penalize accurate predictions. Since the ablation in Table 2 (groups 6 vs 7) attributes a large improvement to the direction component A, the paper must state which sign is actually used in training and correct the equations accordingly.","section":"3.4, Eqs. (16)-(18)"},{"comment":"The fusion rule in Eq. (8) adds matrices of incompatible shapes: W_U^m is an Nn x Nn neighbor self-attention matrix from Eq. (3), while W_S is the target-to-neighbor attention matrix from Eq. (7), which has one target query against Nn neighbor keys and therefore shape 1 x Nn. Unless an omitted reshaping or broadcasting is specified, the sum is undefined, so the central multi-order fusion formula cannot be implemented as written.","section":"3.1, Eq. (8)"},{"comment":"No error bars, confidence intervals, or multiple-seed results are reported for any experiment. Some headline differences are small (e.g., NuScenes FDE 1.56 vs 1.64, and on ETH the proposed method is actually worse than V2-Net-SC on ADE), so without repeated runs and variance reporting the claimed state-of-the-art status is not statistically supported.","section":"4, Tables 1-2"},{"comment":"The paper's conceptual claim that higher-order intention interactions are captured is not supported by the implementation. The layer computes pairwise self-attention among neighbors via Eq. (3), and the stated decomposition assumption ('higher-order intention interactions within the group of neighbors can be decomposed into first-order intention interactions between individual neighbors') is asserted without justification or an ablation that varies the number of interaction hops or compares against a hypergraph/relational baseline. If the claim is meant as an approximation, it should be validated; otherwise the contribution should be framed as pairwise neighbor interaction modeling rather than higher-order fusion.","section":"3.1, Higher-order intention interaction layer"},{"comment":"The directly related higher-order relational reasoning method of Kim et al. [16] is cited in the introduction but is not included in any comparison table. Given the paper's central higher-order-intention claim, omitting the most relevant baseline makes it difficult to assess the marginal value of the proposed multi-order fusion.","section":"4, Table 1"}],"minor_comments":[{"comment":"Equation (10) and the surrounding text are difficult to parse; the reparameterized latent variable equation mixes the definition and the conditional distribution in one line. Please rewrite with a clear definition of q_phi and the sampling operation.","section":"3.2, Eq. (10)"},{"comment":"Equation (18) contains a typographical artifact 'TFP' before the summation; it should be a summation symbol over the prediction horizon.","section":"3.4, Eq. (18)"},{"comment":"The reported NuScenes FDE improvement is inconsistent with Table 1: the text states a 1.87% reduction, but Table 1 lists 1.56 vs 1.64 (4.9%) if compared to E-V2-Net-SC or 1.56 vs 1.59 (1.9%) if compared to V2-Net-SC. Please reconcile the text and the table.","section":"4.1, NuScenes results"},{"comment":"The NLL metric is mentioned in Section 4 as a reported metric, but all NLL results are deferred to a Supplementary Material that is not included with the submission; either include the results or explicitly state that they are omitted.","section":"4, Metrics"},{"comment":"In Eq. (17), the direction vectors are written as \\hat g_i^t - \\hat g_i^{t+1} rather than the standard \\hat g_i^{t+1} - \\hat g_i^t; although the predicted and ground-truth vectors are both reversed so the angle is unchanged, the notation is confusing and should be corrected.","section":"3.4, Eq. (17)"}],"recommendation":"major_revision","confidential_remarks":"The sign inconsistency in Eqs. (16)-(18) and the shape mismatch in Eq. (8) are the main technical blockers. If the authors can confirm that the code uses the corrected signs and can clarify the fusion dimensions, the empirical contribution may be publishable after revision. The missing comparison with Kim et al. [16] and the absence of error bars are also important for the SOTA claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Potential SOTA results on four benchmarks, but the direction-loss sign error is real and load-bearing. Eq. 17 defines Langle = -arccos(cosine similarity). Minimizing Eq. 18 would push predicted directions away from the true direction; maximizing it would conflict with the positive distance error Ldis. The ablation credits A (direction term) with big gains on SDD and NuScenes, so the main empirical claim depends on a sign that is wrong in the paper. As written, the objective is incoherent.\n\nWhat's good: the integration is competent — attention among neighbors, a conditional VAE for trajectory distribution, a KAN-based global refiner, and a direction-sensitive loss. The experimental coverage is broad (ETH/UCY, NBA, SDD, NuScenes) and the reported improvements are double-digit relative gains on several datasets. The ablation is detailed, and code is promised. That is above the usual bar for this subfield.\n\nSoft spots, in order of severity: (1) the sign error above; without a corrected statement we cannot reproduce the loss. (2) No error bars or multiple seeds anywhere; the tables list point estimates only. (3) The 'higher-order intention' framing is oversold. The implementation is multi-head self-attention among neighbors — a pairwise mechanism. The assumption in Section 3.1 that higher-order interactions decompose into first-order ones is asserted, never derived or ablated. The ablation of component I shows the fusion block helps, but doesn't isolate higher-order effects. (4) The KAN novelty is minor; using a recent function approximator as a trajectory refiner is a natural adaptation, not a new concept. (5) The 'squeeze theorem' description is hand-wavy but harmless.\n\nMy read: the empirical numbers, if reproducible, put this at or near SOTA on several benchmarks. But the paper as written cannot be trusted until the loss sign is fixed, the code is released and checked, and variance is reported. I would not cite the numbers yet. This deserves a serious referee: the task is important, the results are interesting, and the integration is solid. A good review would ask for a corrected loss, multi-seed experiments, and a rewrite of the higher-order claims. I would send it to review rather than desk-reject, but with the expectation of major revision.","headline":"Strong reported results on multiple benchmarks, but a load-bearing sign error in the direction loss makes the paper not reproducible as written; needs correction and variance reporting.","tokens_in":14275,"tokens_out":5825,"would_cite":false,"duration_ms":52598,"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":"SocialMOIF claims that fusing direct and indirect neighbor intentions yields state-of-the-art pedestrian trajectory predictions across ETH/UCY, NBA, SDD, and NuScenes.","keywords":["pedestrian trajectory prediction","multi-order intention fusion","higher-order social interaction","trajectory distribution approximator","Kolmogorov-Arnold Networks","distance-direction loss","ADE/FDE","generative trajectory models"],"falsifier":"Run the identical SocialMOIF pipeline on ETH/UCY, NBA, SDD, and NuScenes with the higher-order neighbor-neighbor attention layer removed or replaced by a same-capacity non-interacting mixer, and compare best-of-20 ADE/FDE; if the numbers stay within noise, the claimed higher-order intention mechanism is not what drives the reported gains.","tokens_in":13274,"feed_emoji":"🚶","tokens_out":10735,"duration_ms":91172,"temperature":0.7,"pith_summary":"The paper tries to establish that pedestrian trajectory prediction improves when a model distinguishes two kinds of social intention: the direct influence of each neighbor on the target pedestrian and the indirect influence that neighbors exert on one another before reaching the target. To capture both, the authors build SocialMOIF, which fuses a first-order intention layer with a higher-order intention layer and then uses the fused intention to guide a latent trajectory distribution, refines predictions with a Kolmogorov-Arnold Network optimizer, and trains with a distance-direction loss. If the reported benchmark numbers hold, the model would be the new state of the art on ETH/UCY, the NBA Rebound and Scores subsets, SDD, and NuScenes in best-of-20 average and final displacement error, bettering the previous best ADE/FDE on every one of those benchmarks. The paper's central contribution is a concrete architecture for including indirect group influence alongside direct neighbor interaction without letting higher-order effects overwhelm first-order ones.","feed_headline":"New trajectory model beats previous best on four pedestrian datasets","feed_subtitle":"It fuses direct and indirect neighbor intentions and reports lower ADE/FDE than prior SOTA on all tested benchmarks.","key_machinery":"The load-bearing object is the Multi-Order Intention Fusion (MOIF) module. Its first-order layer computes an attention matrix between the target and each neighbor; its higher-order layer computes a multi-head self-attention matrix among the neighbors themselves, which the paper justifies by decomposing group intentions into pairwise first-order interactions. The two matrices are combined by adding the higher-order matrices scaled by learned influence factors to the first-order matrix, then using the result to attend to value vectors. The fused intention feeds a VAE-style trajectory distribution approximator, where an RNN updates the latent variable and a future-trajectory embedding serves as an upper bound; the decoder output is then refined by a KAN-based global trajectory optimizer operating over the entire time dimension at once, and the full system is supervised by the sum of displacement error and angle error.","core_discovery":"The paper's central discovery is that a multi-order intention fusion layer—one head for direct target-neighbor interactions, one head for neighbor-neighbor interactions within the group, combined with per-subspace learned weights—can reduce prediction errors across datasets that differ in crowd density, motion purpose, and static or dynamic scenes. In the paper's tables, best-of-20 ADE/FDE drop from 0.15/0.20 to 0.13/0.18 on ETH/UCY, from 0.54/0.79 to 0.34/0.66 on NBA Rebound, from 0.46/0.76 to 0.30/0.56 on NBA Scores, from 0.21/0.34 to 0.17/0.24 on SDD, and from 1.04/1.64 to 0.92/1.56 on NuScenes. The paper attributes these gains to four interacting components: the Multi-Order Intention Fusion module, a trajectory distribution approximator that explicitly models the latent intent distribution during training, a global optimizer built from Kolmogorov-Arnold Networks that predicts all future time steps in parallel, and a loss function that penalizes both distance error and direction mismatch.","pith_inferences":["Beyond the paper, because the higher-order layer is mathematically a set of pairwise attention operations, the reported gains could come from extra model capacity rather than from true group-level reasoning; swapping that layer for a same-capacity non-interacting mixer would settle this.","Beyond the paper, the 'squeeze theorem' description is an analogy unless one proves that the sampled latent distribution is actually bounded by the fused intention and the future-trajectory embedding; checking that bound empirically would strengthen the interpretability claim.","Beyond the paper, the largest relative ADE gain in the table is on NBA Rebound, a strongly goal-directed dataset, which suggests the approach may show its clearest value in sports and other purposeful-motion settings, and long-horizon team-sport prediction would be a natural stress test.","Beyond the paper, best-of-20 ADE/FDE can reward a wide proposal distribution rather than per-mode accuracy; reporting best-of-1 and per-sample NLL alongside best-of-20 would show whether the improvement is genuine accuracy or mostly diversity."],"forward_implications":["If the reported results reproduce, SocialMOIF becomes the best-published best-of-20 ADE/FDE model on ETH/UCY, NBA Rebound, NBA Scores, SDD, and NuScenes at the time of the work.","The explicit latent-variable updating scheme means the training-time trajectory distribution is directly supervised rather than left implicit, which the paper argues improves interpretability.","Predicting all future frames in parallel through the KAN-based optimizer removes sequential decoding dependency, reducing error accumulation and enabling faster inference.","The directional loss term is claimed to improve other models as well when added, making it a separable training recipe independent of the rest of SocialMOIF."],"supporting_citations":[{"why":"Prior work that modeled 1st-to-N-order intentions and is the approach this paper says overweights higher orders, motivating the multi-order fusion design.","marker":"[16]"},{"why":"Cited for the reparameterization technique used to sample latent variables in the trajectory distribution approximator.","marker":"[17]"},{"why":"Introduces Kolmogorov-Arnold Networks, which the paper adopts in the global trajectory optimizer.","marker":"[23]"},{"why":"Second KAN formulation cited for the optimizer's activation-function representation.","marker":"[24]"},{"why":"E-V2-Net-SC is the previous best on ETH/UCY, NBA, and NuScenes in the paper's comparison, so its ADE/FDE values are the numbers to beat.","marker":"[40]"},{"why":"Provides the angle-based social interaction representation behind the V2-Net-SC family of baselines the paper improves on.","marker":"[41]"},{"why":"EqMotion is a strong baseline across all four benchmark groups, providing the equivariant multi-agent comparison.","marker":"[46]"},{"why":"Neural social physics baseline that the paper says held the best SDD result before this work.","marker":"[52]"}],"fun_headline_variants":["SocialMOIF fuses direct and indirect neighbor intents for sharper trajectory forecasts","Multi-order intention fusion cuts pedestrian prediction error on five benchmarks","Higher-order intent fusion improves trajectory accuracy across static and dynamic scenes","SocialMOIF's multi-order fusion reduces ADE and FDE on all tested datasets","Direct plus indirect neighbor intentions key to better pedestrian forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's higher-order layer is built from pairwise self-attention among neighbors, so the central claim depends on group-level effects on a pedestrian being fully decomposable into pairwise neighbor-to-neighbor effects.","fun_headline_variants_meta":{"raw":{"variants":["SocialMOIF fuses direct and indirect neighbor intents for sharper trajectory forecasts","Multi-order intention fusion cuts pedestrian prediction error on five benchmarks","Higher-order intent fusion improves trajectory accuracy across static and dynamic scenes","SocialMOIF's multi-order fusion reduces ADE and FDE on all tested datasets","Direct plus indirect neighbor intentions key to better pedestrian forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001232,"raw_usage":{"total_tokens":5089,"prompt_tokens":1003,"completion_tokens":4086,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":3995}},"tokens_in":619,"tokens_out":4086,"duration_ms":23829,"temperature":1.0,"reasoning_tokens":3995,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:21:33.136273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical SocialMOIF pipeline on ETH/UCY, NBA, SDD, and NuScenes with the higher-order neighbor-neighbor attention layer removed or replaced by a same-capacity non-interacting mixer, and compare best-of-20 ADE/FDE; if the numbers stay within noise, the claimed higher-order intention mechanism is not what drives the reported gains.","supporting_citations":[{"cited_title":"Machine learn- ing for emergency management: A survey and future outlook","cited_arxiv_id":null,"evidence_quote":"Cited for the reparameterization technique used to sample latent variables in the trajectory distribution approximator."},{"cited_title":"Socialcircle: Learning the angle-based social in- teraction representation for pedestrian trajectory prediction","cited_arxiv_id":null,"evidence_quote":"Provides the angle-based social interaction representation behind the V2-Net-SC family of baselines the paper improves on."},{"cited_title":"Human tra- jectory prediction via neural social physics","cited_arxiv_id":null,"evidence_quote":"Neural social physics baseline that the paper says held the best SDD result before this work."}],"review_version":1}