{"id":"61d0c08e-d138-4ea2-a672-8c19304e4c7c","arxiv_id":"2608.03379","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A prior-guided residual flow matching framework with dynamic cross-interaction reports state-of-the-art results on 3D multi-person motion forecasting benchmarks.","lead":"A new model for predicting how several people will move in 3D first makes a rough guess, then refines it using a generative model over the difference between that guess and the real motion. It reports lower errors than previous methods on several public benchmarks, and its main twist is a time-varying gate that controls how strongly people influence each other during refinement.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main text describes single-step decoding, but supplementary and conclusion describe 10-step ODE integration; this contradiction determines whether DCI is active at inference, so the reported SOTA cannot be attributed to the described mechanism.","rationale":"The reader's verdict of REJECT is supported by the most load-bearing concern: the inference procedure is described inconsistently, and the discrepancy determines whether the central DCI mechanism is actually exercised at inference. This is not a stylistic quibble; it directly affects the interpretation of every experimental result and the attribution of the reported SOTA to the proposed components. The reader's weakest_assumption focused on the DCI gate schedule being a capacity artifact rather than a noise-suppression effect; that is a legitimate secondary concern, but the more fundamental problem is that, depending on which inference description is true, the DCI mechanism either does nothing at inference or was evaluated under an iterative procedure that the main text explicitly denies. This ambiguity is enough to reject the paper as written, even though the experimental numbers could in principle be reproduced if the authors clarify and release code. The absence of error bars and the use of custom Mix1/Mix2 splits further weaken the empirical claim, but they are secondary to the internal contradiction. I agree with the reader's overall rejection, though I would ground it primarily in the inference inconsistency rather than the DCI gate's theoretical justification.","tokens_in":19206,"tokens_out":4174,"duration_ms":45270,"concrete_test":"Inspect the official repository's inference entry point (e.g., test.py or predict.py). Count the number of ODE solver steps executed. If the code runs 10 sequential Euler updates over t∈[0,1), then the main text's 'single step' statement is contradicted; re-run the w/o DCI ablation under the same 10-step solver and check whether the reported 3DPW-RC MPJPE gap (70.62→69.92) persists. If the code decodes K endpoints in one forward pass with no integration loop, then the DCI gate g(t) is evaluated at a single t; re-run the w/o DCI ablation with g fixed to the value at that t and verify whether the performance gap remains.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central methodological claim is internally contradictory about inference. Section 'Future Motion Prediction' states: 'Leveraging the endpoint parameterization, K parallel prediction heads ρk directly decode distinct multi-modal residuals in a single step' (Eqs. 11–12). However, the supplementary 'Flow Construction and ODE Sampling' section and the Conclusion state that inference solves an ODE with a fixed-step explicit Euler solver using 10 integration steps over t∈[0,1), with the update in Eq. 19. These two descriptions cannot both be true. If inference is truly single-step, then the DCI gate g(t)=0.6+0.4t^1.5 (Eq. 9) is evaluated at a single t (likely t=1), so the 'Dynamic Cross-Interaction' mechanism—the paper's second main contribution and the subject of its w/ vs w/o DCI ablation—has no dynamic role at inference. The time-varying modulation would be a training-time artifact, and the claimed benefit of 'synchronizing spatial message-passing with integration progress' would not apply to the deployed model. If inference is actually the 10-step ODE, then the main-text description is false, and the reported experimental numbers correspond to an iterative procedure whose latency is acknowledged in the Conclusion. Either way, the empirical SOTA claim is not attributable to a consistently described method. A further unresolved discrepancy: the abstract states code is available at a GitHub URL, while the supplementary says code 'will be fully released to the public once this paper is accepted.' Together, these inconsistencies make the central claim—that the proposed framework with DCI achieves state-of-the-art accuracy—impossible to verify from the manuscript as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for 3D multi-person motion prediction that combines a Deterministic Coarse Prior (DCP) with a conditional flow-matching model over motion residuals, plus a Dynamic Cross-Interaction (DCI) gate that scales cross-person message-passing with the integration progress t. The authors report state-of-the-art results on 3DPW, CMU-Syn, Mix1/Mix2, and cross-dataset settings, with ablations attributed to DCP, residual flow, DCI, joint branch, and full-sequence temporal alignment. The central methodological description, however, is internally inconsistent: the main text describes single-step decoding from K parallel heads, while the supplementary material and the conclusion describe an iterative 10-step Euler ODE solver at inference.","tokens_in":19550,"tokens_out":4467,"duration_ms":50100,"significance":"If the reported results are reproducible and the inference procedure is unambiguously specified, the framework would be a strong empirical contribution: it extends recent EMPMP-style hierarchical motion modeling with a residual flow objective and a time-gated social interaction module, and it is evaluated across multiple benchmarks and metrics. The paper also provides extensive ablations and qualitative comparisons, which strengthens the empirical case. However, the contribution currently cannot be cleanly assessed because the main text and the supplementary material disagree on a load-bearing implementation detail—whether DCI is active over a multi-step ODE trajectory or is evaluated in a single decoding step. That inconsistency must be resolved before the SOTA claim can be attributed to the described mechanism. The DCI ablation also conflates temporal modulation with overall attenuation, so the mechanism's specific benefit is not yet isolated.","major_comments":[{"comment":"The main text states that the K parallel prediction heads decode residuals in a single step, but the supplementary material describes a fixed-step explicit Euler solver with 10 integration steps over t in [0,1), and the Conclusion explicitly acknowledges that 'iterative ODE integration in the flow matching process incurs higher inference latency.' These two descriptions are mutually exclusive. If inference is the 10-step ODE, the main text is false; if inference is single-step, then the DCI gate g(t)=0.6+0.4 t^1.5 in Eq. (9) is applied at only one value of t, so the 'Dynamic Cross-Interaction' mechanism has no dynamic role at inference and the DCI ablation in Table 3 does not test what it claims. This must be corrected and the actual inference procedure, including the number of steps and the value(s) of t used at test time, must be stated unambiguously before the reported SOTA numbers ca","section":"Future Motion Prediction (Eqs. 11-12) vs. Supplementary 'Flow Construction and ODE Sampling' (Eq. 19) vs. Conclusion"},{"comment":"The only DCI ablation replaces the time-varying gate with the static interaction used in the coarse stage, i.e., g(t)=1. This comparison conflates the effect of temporal modulation with the effect of overall gate strength: a constant gate of 1.0 may simply inject more cross-agent information at every step, while a constant gate of 0.6 would test whether attenuation alone, without time-dependence, is beneficial. The premise that early flow states are noise-dominated and that the monotone power-law schedule is the correct temporal profile is asserted rather than derived. The supplementary hyperparameter analysis (wmin, wmax, gamma) is useful but still does not include a time-invariant gate at matched average strength. Please add constant-gate baselines at g=0.6 and g=1.0, and, if possible, a direct diagnostic of how prediction error varies with the integration step under different gates.","section":"Ablation Study: 'Effectiveness of Dynamic Cross-Interaction' (Table 3)"},{"comment":"The abstract states that code is available at a GitHub URL, while the supplementary material states that code and trained models 'will be fully released to the public once this paper is accepted.' These statements are contradictory. Please harmonize the availability statement and, if the URL is indeed live, verify that it contains the code; if not, remove the URL or mark it as anonymous for review. This is a presentation issue but it affects reproducibility claims.","section":"Abstract vs. Supplementary Material (Code Availability)"}],"minor_comments":[{"comment":"The scaling parameter for the noise term is called beta in Eq. (7) and sigma in Eq. (18); both are later reported as 2.0. Unify the notation to avoid confusion.","section":"Eq. (7) / Eq. (18)"},{"comment":"The abbreviation 'AFM' in the row 'w/ AFM' is not expanded or explained anywhere in the text. It appears to mean 'Absolute Flow Matching,' but it should be defined at first use.","section":"Ablation Study (Table 3)"},{"comment":"The column header of Table 1 is garbled: the grouping 'Ori 16f/14f RC 16f/14f 2s/2s 1s/1s' is repeated, and the column labels do not clearly map to the reported metrics. Please reformat the table so that each dataset, protocol, and metric is unambiguously labeled.","section":"Table 1"},{"comment":"The caption says the results are for the '1s-input/3s-output setting,' but the column headers read 'Out Length 1s 2s 3s.' Clarify whether these columns are three output horizons within a single protocol or three separate settings.","section":"Table 2"},{"comment":"The sentence 'To accommodate asymmetric lengths, we define a unified sequence length I=max(T_h, T_f)' is followed by a procedure for padding/truncation; the wording is slightly confusing because the DCT length is I but the coordinate conversion and truncation are only described afterward. Consider rewriting for clarity.","section":"Supplementary 'Details on Network and Training'"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern lands: the mismatch between single-step decoding and 10-step ODE sampling is not a cosmetic issue—it determines whether DCI is actually dynamic at inference. I would ask the authors to (1) state exactly how inference is performed, (2) adjust the main-text description accordingly, and (3) add a constant-gate DCI ablation. If the authors instead argue that the single-step description is the intended one, then the DCI claims need substantial rewriting. The empirical comparisons are extensive and the method is plausible, so a major revision rather than a rejection seems appropriate at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll give you my honest read of arXiv:2608.03379. The core idea is actually reasonable: instead of generating multi-person skeletons from pure noise, they first predict a deterministic coarse prior (DCP), then run a flow matching model on residuals. That's a worthwhile decomposition, and the ablations show each component—DCP, residual flow over absolute flow, the joint branch, and the dynamic cross-interaction (DCI)—gives a consistent improvement. The gains over EMPMP across 3DPW, CMU-Syn, and Mix1/Mix2 are broad, and the tables are clean. I'll credit the empirical discipline in the ablation study, even though the architecture closely follows EMPMP.\n\nThe problem is that the paper describes two incompatible inference procedures. The main text (Section 'Future Motion Prediction', Eq. 12) says K parallel heads decode residuals in a single step. The supplementary (Eq. 19) and the Conclusion describe a 10-step explicit Euler ODE integration with t in [0,1). These cannot both be true. If inference is really the 10-step ODE, the main text is false; if it is single-step, then DCI's time-varying gate g(t)=0.6+0.4t^1.5 is evaluated at one point, and the mechanism's stated purpose—synchronizing message passing with integration progress—is vacuous at inference. The w/o DCI ablation does not resolve this; replacing the gate with a constant would behave the same under either interpretation. So the central methodological claim is not attributable to a consistently described model.\n\nOther issues: the abstract says code is available at a GitHub URL, while the supplementary says code will be released once accepted; no error bars are given; Mix1/Mix2 use custom train/test splits, so comparisons against published numbers depend on baseline reproductions; and the DCI hyperparameters (w_min, w_max, gamma) are tuned on validation data. None of these alone is fatal, but they add up when the core inference is contradictory.\n\nFor whom is this paper? The target audience is multi-person motion forecasting researchers who would care about the residual flow matching idea and the DCI gating schedule. The empirical SOTA claim, if real, is a useful incremental advance. But as written, I cannot recommend citing it or using its numbers. I would send it back for major revision, asking the authors to fix the inference description, release code and exact splits, and ideally report error bars. It deserves careful refereeing rather than a desk reject, because the underlying approach is sound and the ablations are well-formed.\n\nRecommendation: engage with it, but demand the consistency fix before accepting.","headline":"A sensible two-stage residual flow matching decomposition with a real contribution, but a load-bearing contradiction between the main text and supplementary about single-step vs. 10-step ODE inference makes the reported SOTA claims impossible to verify as written.","tokens_in":20081,"tokens_out":4397,"would_cite":false,"duration_ms":46714,"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":"The paper claims that residual flow matching with a deterministic prior and dynamic cross-interaction achieves state-of-the-art accuracy in 3D multi-person motion prediction.","keywords":["3D multi-person motion prediction","flow matching","residual generative models","dynamic cross-interaction","ODE integration scheduling","multi-hypothesis forecasting","social interaction modeling","motion forecasting"],"falsifier":"Retrain with the gate reversed, $g(t)=0.6+0.4(1-t)^{1.5}$, keeping everything else identical; if MPJPE on 3DPW-RC stays comparable or improves, the benefit is not from keeping noisy early states local.","tokens_in":19073,"feed_emoji":"🏃","tokens_out":12806,"duration_ms":137149,"temperature":0.7,"pith_summary":"The paper is trying to establish that 3D multi-person motion prediction is better done as a two-stage generative problem than as either a deterministic forecast or a pure-noise generative flow. It anchors generation with a deterministic coarse prior, runs flow matching over motion residuals instead of raw skeletons, and gates cross-person attention by how far the ODE integration has progressed. If the claim holds, the practical payoff is more accurate, structurally plausible multi-person forecasts in crowded scenes and across dataset shifts. The paper reports lower errors than the compared methods on five metrics over 3DPW, CMU-Syn, Mix1/Mix2, and cross-dataset transfer to MuPoTS-3D.","feed_headline":"Residual flow beats prior models on multi-person 3D motion prediction","feed_subtitle":"A coarse prior keeps skeletons stable while a time gate restores social cues only when generated states are clean enough to trust.","key_machinery":"The load-bearing object is the residual flow state $R^{(t)}=tR^{(1)}+(1-t)\\beta R^{(0)}$, with the clean endpoint defined as $R^{(1)}=Y-X_{T_h}$, so the network predicts the residual from the last observed pose rather than the absolute skeleton. The second mechanism is the Dynamic Cross-Interaction gate $g(t)=w_{\\min}+(w_{\\max}-w_{\\min})t^\\gamma$, instantiated as $0.6+0.4t^{1.5}$, which multiplies the global-to-local message-passing terms in each refinement layer. The deterministic spatial prior $D_c$ stays unattenuated while the uncertain global context is gated, which is what suppresses early-stage noise interference.","core_discovery":"The authors claim that the hard part of multi-person motion prediction is not sampling diverse futures but keeping those futures kinematically coherent while agents interact. Their decomposition lets a Deterministic Coarse Prior produce a first future, then a Residual Flow Matching module learns a conditional flow over the residual between that prior and the true future, so the generative model never has to create skeletons from isotropic noise. A Dynamic Cross-Interaction mechanism scales cross-person feature injection by $g(t)=0.6+0.4t^{1.5}$ over the ODE trajectory, leaving early noise-dominated states largely local and restoring social context as $t\\to1$. With Winner-Takes-All endpoint t","pith_inferences":["Editorial extension: the gate is hand-set, so a direct next test is to learn $g(t)$ from the noise level of the intermediate state; a learned schedule could outperform the fixed power law on horizons where early states are cleaner or noisier.","Editorial extension: the coarse-anchor-plus-residual-flow split is broader than skeletons; the same recipe could reduce generative load in trajectory forecasting where a cheap deterministic planner provides the anchor.","Editorial extension: the paper's own stated limitations—ODE latency, preset gate, and fixed person count—bound the claim and identify where a follow-up would need to generalize it, not where the current evidence is silent.","Editorial extension: measuring the signal-to-noise ratio of $R^{(t)}$ at small $t$ would test the stated motivation directly; if early states are not noise-dominated, the gate's benefit would be better explained as a capacity rebalancing effect."],"forward_implications":["A single training recipe can combine deterministic structure with generative diversity: the coarse prior supplies the skeleton, the residual flow supplies the multiple plausible futures.","Because the flow is parameterized by the clean endpoint, a short 10-step Euler solver is enough at inference, making the accuracy gain compatible with practical use.","The gating principle implies that interaction strength should track the reliability of intermediate states, not just the time index, which is a general lesson for multi-agent generative models.","Reported cross-dataset results suggest the learned interaction representations transfer from synthetic to real scenes rather than memorizing the training domain.","The model's accuracy in the 10-person Mix2 setting suggests that dense social context can be used productively rather than treated as extra noise."],"supporting_citations":[{"why":"Introduces flow matching, the generative formulation the paper re-centres on motion residuals.","marker":"Lipman et al. 2023"},{"why":"Supplies the endpoint parameterization and the flow-based trajectory forecasting practice adopted by the residual refinement stage.","marker":"Fu et al. 2025"},{"why":"Source of the DCT-based temporal encoding, the local-global decoupling design, and the EMPMP baseline and evaluation protocol.","marker":"Zheng, Yu, and Sun 2025"},{"why":"The JRT baseline for interaction-aware multi-person forecasting that the method must outperform.","marker":"Xu et al. 2023"},{"why":"Defines the Mix1/Mix2 high-density benchmarks and the TBIFormer synthesis protocol used for crowded-scene evaluation.","marker":"Peng, Mao, and Wu 2023"},{"why":"SocialDiffusion exemplifies the uniform cross-person aggregation during denoising that the Dynamic Cross-Interaction mechanism replaces.","marker":"Tanke et al. 2023"},{"why":"Establishes timestep-aware conditioning through PhysDiff, the precedent DCI extends to inter-agent interaction gating.","marker":"Yuan et al. 2023"},{"why":"T2P baseline and source of the Best-of-K protocol used for multi-hypothesis evaluation.","marker":"Jeong, Park, and Yoon 2024"},{"why":"Provides the 3DPW in-the-wild dataset used for the main real-world evaluation.","marker":"von Marcard et al. 2018"},{"why":"Provides the CMU-Syn synthetic multi-person benchmark and the multi-range transformer baseline.","marker":"Wang et al. 2021"}],"fun_headline_variants":["Residual flow keeps multi-person motion coherent","Prior-guided residual flow shapes social motion","Time-gated social cues stabilize motion forecast","Motion prediction via residual flow and timed interactions","Dynamic gating improves multi-person motion prediction"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that early ODE integration states are so corrupted by noise that suppressing cross-person attention according to a hand-set schedule is what creates the accuracy gain; if that premise is false, the reported improvement could be an artifact of changing model capacity instead of evidence about noise.","fun_headline_variants_meta":{"raw":{"variants":["Residual flow keeps multi-person motion coherent","Prior-guided residual flow shapes social motion","Time-gated social cues stabilize motion forecast","Motion prediction via residual flow and timed interactions","Dynamic gating improves multi-person motion prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00013,"raw_usage":{"total_tokens":957,"prompt_tokens":731,"completion_tokens":226,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":175}},"tokens_in":475,"tokens_out":226,"duration_ms":3090,"temperature":1.0,"reasoning_tokens":175,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:01:19.032719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain with the gate reversed, $g(t)=0.6+0.4(1-t)^{1.5}$, keeping everything else identical; if MPJPE on 3DPW-RC stays comparable or improves, the benefit is not from keeping noisy early states local.","supporting_citations":[],"review_version":1}