{"id":"6f9f2515-b620-4ff0-a61e-e478854faf93","arxiv_id":"2507.07515","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GGMotion, a grouped equivariant graph network with spatio-temporal radial fields and a parallel dynamics-kinematics update, reports the lowest average MPJPE among compared baselines on Human3.6M, CMU-Mocap, and 3DPW for short-term prediction.","lead":"A new neural network called GGMotion predicts future human body poses by grouping joints and using geometry-aware graph layers that respect rotations and translations. It reports small but consistent accuracy gains over earlier methods on three standard 3D motion benchmarks, especially for short-term predictions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (14)'s L_aux shrinks predicted bone lengths toward zero; it is a smoothing regularizer, not the advertised joint-length constraint, so the method as described is not what Table IX ablates.","rationale":"The paper's central claim is that GGMotion outperforms state-of-the-art baselines on human motion prediction using a lightweight equivariant model with a dynamics-kinematics prior enforced through an auxiliary loss. The most concrete place where the described mechanism fails is Section III-E: Eq. (14) defines L_aux as the L1 norm of each predicted bone vector, with no reference length to compare against. Minimizing that term drives predicted bone lengths toward zero, which is not a kinematic constraint on joint lengths. Table IX attributes consistent improvements to this term, so the empirical support for the loss is tied to an objective that the text does not correctly describe. This is an internal inconsistency between the stated purpose and the formula, not a disagreement with community consensus. A single check of the released code can distinguish between a typo in the paper and a mismatch in the actual training objective. I agree with the reader's weakest assumption, and I do not think this requires changing the conditional verdict: the concern justifies a request for code inspection and a corrected formula, but it does not by itself establish that the benchmark numbers are wrong. The authors should also report multiple seeds, because the reported margins over EqMotion and KSOF are only 0.1-1.1 mm on average, and the ablation differences are even smaller; without variance estimates, the statistical significance of the headline comparison is unresolved.","tokens_in":19937,"tokens_out":10308,"duration_ms":111579,"concrete_test":"Inspect the released training code at https://github.com/inkcat520/GGMotion.git and locate the computation of L_aux. Check whether it computes ||yhat_i - yhat_j||_1 exactly as in Eq. (14) or whether it compares predicted bone lengths to a reference length such as ||x_i - x_j|| from the initial/ground-truth pose. If the implementation matches Eq. (14), implement a true rest-length loss and rerun the Table IX ablation; if the improvement disappears, the reported gain is caused by a different mechanism than the one claimed. If the implementation differs from Eq. (14), rerun the ablation using the formula printed in the paper to determine whether the published results correspond to the published method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-E defines L_aux as the mean L1 norm of each predicted bone vector: sum over t and over adjacent joints i,j of ||yhat_i^t - yhat_j^t||_1. No ground-truth bone length or initial rest length appears anywhere in the formula. Minimizing this term pulls every adjacent pair of predicted joints together, which is a shrinkage/smoothing regularizer, not a constraint enforcing fixed skeletal lengths. The text explicitly says the loss 'introduce[s] joint length as an auxiliary loss' and Table IX credits this term with consistent improvements, so the stated purpose and the objective are in direct conflict. If the released code matches Eq. (14), the claimed kinematic-constraint mechanism is absent and the reported gains come from an undocumented regularizer. If the code instead compares predicted bone lengths to reference lengths, then the paper's methodological description is inaccurate and the equation must be corrected. Either way, the link between the central contribution and the experiments is broken. This matters because the reported margins over the closest baselines are only 0.1-1.1 mm on average, and no error bars are provided, so the ablation result is the main evidence that the auxiliary loss contributes to the headline performance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GGMotion, a lightweight group-graph network for 3D human motion prediction. It combines a spatio-temporal radial field, inter-group/intra-group interaction modules, equivariant MLPs, and a parallel dynamics-kinematics update, and claims state-of-the-art short-term MPJPE on Human3.6M, CMU-Mocap, and 3DPW with only 0.46M parameters.","tokens_in":20184,"tokens_out":5535,"duration_ms":56947,"significance":"If the reported numbers are accurate, GGMotion would be a competitive lightweight equivariant model for short-term motion prediction, and the grouping strategy is a reasonable inductive bias. The paper offers a public code release, evaluates on three standard benchmarks, and includes ablations for each module, which is commendable. However, the auxiliary loss is not what it is claimed to be, and the equivariance proof is too terse to verify, so the central claims currently require revision before the results can be fully trusted.","major_comments":[{"comment":"The auxiliary loss L_aux as written in the unnumbered equation in Section III-E is the mean L1 distance between predicted positions of adjacent joints; it contains no reference to ground-truth bone lengths or rest lengths. Minimizing this sum pulls predicted joints together, so it is a shrinkage/smoothing regularizer rather than the advertised 'joint length' constraint. The text explicitly states that the loss 'introduce[s] joint length as an auxiliary loss' and Table IX credits this loss with consistent improvements, creating a direct contradiction between the stated purpose and the objective. If the released code implements a different comparison against reference lengths, Eq. (14) must be corrected; if the code matches the formula, the reported gains come from an undocumented regularizer. Please clarify and update the description and the ablation discussion accordingly.","section":"Section III-E, 'Loss Function'"},{"comment":"The claim that phi_eq is O(3)-equivariant is not demonstrated. The covariance matrix Sigma_n = Z_q^T Z_k is invariant under rotations only if the learned linear layers W_q, W_k are equivariant maps or act solely on the feature dimension, but the manuscript does not specify their structure or show how R acts on the concatenated variables. The sentence 'after subtracting the centroid ... we can easily conclude phi_eq exhibits equivariance' is too terse for a load-bearing property; please provide a step-by-step proof or state the exact constraints on W_q, W_k, W_v and the MLP. Without this, the claimed benefit of 'preserving geometric equivariance' is not established.","section":"Section III-C, 'Equivariant MLP'"}],"minor_comments":[{"comment":"In the ablation text following Table VI, the sentence 'the T. field achieved lower prediction errors than the T. field' appears to have a typo; the second reference should presumably be 'S. field'. Please correct it.","section":"Section V, Table VI"},{"comment":"The reported improvements over the closest baselines are often below 1 mm (e.g., 9.0 vs 9.3 mm at 80 ms and 45.1 vs 45.8 mm at 320 ms on Human3.6M) and no variance or significance information is given; please report mean and standard deviation over multiple runs or seeds so the reader can judge the margin.","section":"Section IV-D, Table I"},{"comment":"The dynamics update in Eq. (8) is a learned MLP over [f_j, r_ij, v_ij]; calling it the 'dynamics-kinematics algorithm' may overstate its connection to the hand-derived Eq. (1). Suggest rewording to 'learned dynamics-style update' or clarifying that it is parametric.","section":"Section III-D, Eq. (8)"},{"comment":"The contribution bullet describing the 'equivariant MLP integrated with the self-attention mechanism' would benefit from explicitly contrasting the covariance-based weighting with standard softmax attention, since the current text uses 'attention weight computation' only loosely.","section":"Section I, contributions"}],"recommendation":"major_revision","confidential_remarks":"The main blocker is the inconsistency between the auxiliary loss formula and its description; the equivariance proof also needs tightening. If the code matches the formula, the loss is a shrinkage regularizer, which may still be a legitimate trick but should be described accurately. I recommend asking the authors to provide the code or correct Eq. (14) and to expand the equivariance proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The architecture is a real, coherent combination—spatio-temporal radial field with centroid edges, six-group body partition, covariance-based equivariant attention, and a parallel 'dynamics-kinematics' update—and it reports consistent, if modest, gains on three benchmarks with a small model. Second: the auxiliary loss in Eq. (14) does not do what the text says, and this needs fixing before the paper can be trusted.\n\nWhat is new: prior equivariant motion models aggregate spatial neighbor features only; GGMotion adds temporal edges to the centroid with learnable scaling, groups the skeleton to learn per-body-part dynamics, and uses covariance-matrix attention inside an equivariant MLP. The ablations are reasonably thorough: modules, number of blocks, group count, and loss terms. The reported improvements are small—around 0.1–1.1 mm average MPJPE at 400 ms—but consistent across Human3.6M, CMU, and 3DPW. If the implementation matches the description, this is a useful incremental contribution.\n\nThe soft spots are real but localized. Eq. (14) defines L_aux as the mean L1 norm of each predicted bone vector, with no ground-truth or rest length anywhere. Minimizing that pulls adjacent predicted joints together; it is a shrinkage/smoothing regularizer, not a joint-length constraint. The text and Table IX credit this term with consistent gains, so either the formula is wrong or the stated mechanism is absent. That is a reproducibility problem, and it is the main reason I would not accept the paper as-is. Related: no error bars are reported, and the margins over strong baselines are thin enough that a few seeds could change the ranking. There is also an internal inconsistency in parameter count—abstract and Table VII say 0.46M, Table V says 0.76M for the same model—and a typo in Table VI where the spatial and temporal ablations are described with the same name.\n\nI do not think the central architecture is flawed. The learned dynamics update is not a derived physical algorithm, but the paper mostly presents it as a parametric approximation, which is acceptable. The main fix is to correct or clarify the auxiliary loss, add error bars, and reconcile the parameter counts.\n\nThis paper deserves a serious referee. It is an incremental but well-constructed contribution to human motion prediction; I would send it to review with a request for major revision on the loss issue and reporting. I would not cite it in my own work until the loss question is resolved.","headline":"A genuinely new equivariant architecture with modest but consistent gains, held back by an auxiliary loss formula that does not match its stated purpose.","tokens_in":20720,"tokens_out":2397,"would_cite":false,"duration_ms":26028,"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":"GGMotion treats the human body as a set of groups updated by parallel dynamics-kinematics rules and reports state-of-the-art short-term motion prediction accuracy on three benchmarks.","keywords":["human motion prediction","equivariant graph neural networks","dynamics and kinematics","spatio-temporal radial fields","group graph networks","skeleton topology","MPJPE"],"falsifier":"Inspect the released code at the paper's repository and run the published $L_{\\text{aux}}$ formula on predicted sequences: if average bone length over the test set decreases monotonically with training while the loss drops, the kinematic-constraint interpretation fails; alternatively, compare predictions trained with $L_{\\text{aux}}$ to ones trained without it and measure whether per-bone variance relative to ground-truth bone lengths actually shrinks.","tokens_in":19686,"feed_emoji":"🦴","tokens_out":6739,"duration_ms":66606,"temperature":0.7,"pith_summary":"GGMotion claims that human motion prediction improves when the skeleton is treated not as one abstract graph but as a set of body groups, each updated by explicit dynamics and kinematics rules that respect Euclidean geometry. The network uses spatio-temporal radial fields to aggregate geometric edge features, and equivariant MLPs keep the representation equivariant under rotations, reflections, and translations. Joint positions are updated in parallel through a rigid-body dynamics step followed by kinematic integration, with inter-group and intra-group interaction modules capturing dependencies at different scales. On Human3.6M, CMU-Mocap, and 3DPW, the method reports the best short-term average MPJPE among compared baselines, for instance 9.0 mm at 80 ms and 45.1 mm at 320 ms on Human3.6M, while stay within a compact parameter budget. The significance would be a lightweight, physically grounded predictor that produces more plausible short-term motion.","feed_headline":"Group-graph physics model tops short-term motion prediction","feed_subtitle":"A compact equivariant network reports 9.0 mm error at 80 ms on Human3.6M.","key_machinery":"The spatio-temporal radial field is the central mechanism: it computes spatial edge weights from Euclidean distances and hop counts between joints, and temporal edge weights from each joint's distance to the body centroid, then combines them with learnable scaling factors to form a motion force. This force is processed by an equivariant MLP whose attention is computed from a covariance matrix rather than a softmax, preserving $\\mathrm{O}(3)$ equivariance. The dynamics-kinematics propagation then updates each group in parallel: Eq. (8) refines acceleration from force, position difference $r_{ij}$, and velocity difference $v_{ij}$, and Eqs. (9)-(10) integrate velocity and position.","core_discovery":"The paper's central claim is that explicit group-wise dynamics-kinematics modeling, combined with a spatio-temporal radial field, yields more accurate and physically plausible human motion prediction than full-body graph networks or simple upper/lower body splits. The model is equivariant under the orthogonal group $\\mathrm{O}(3)$ and updates features through a parallel dynamics step that refines joint accelerations from position difference, velocity difference, and force, followed by kinematic integration of velocity and position. The authors report that this design achieves state-of-the-art average MPJPE on the three benchmarks for short-term horizons, including 9.0 mm at 80 ms and 45.1 mm at 320 ms on Human3.6M, and outperforms the equivariant baseline EqMotion at 1000 ms by a clear margin, which they attribute to the temporal edges of the radial field.","pith_inferences":["The published auxiliary loss formula penalizes the L1 norm of each predicted bone vector with no reference length, so the loss as written would drive bone lengths toward zero; if the released implementation instead compares to reference skeleton lengths, the paper's description and the ablation's interpretation do not match.","Enforcing invariance under reflections (det = -1 rotations) could remove handedness information relevant to actions like turning; a testable variant would restrict equivariance to proper rotations $\\mathrm{SO}(3)$ and compare accuracy.","The fixed six-group partition is tuned for the Human3.6M skeleton; applying it to skeletons with different joint counts may require re-selecting groups, and a data-driven group assignment could transfer better across benchmarks.","The radial field's temporal edge is defined relative to the body centroid; a richer temporal model that also connects joints across time steps might close the remaining gap to DCT/TCN baselines in long-term prediction."],"forward_implications":["Short-term prediction error on Human3.6M drops to 9.0 mm at 80 ms and 45.1 mm at 320 ms, below all compared baselines.","The same network generalizes to CMU-Mocap and 3DPW, improving average MPJPE across short-term horizons on both datasets.","Grouping into six body parts beats treating the whole body as one group or splitting into upper and lower halves, supporting the grouping strategy over fixed topology splits.","The temporal radial field contributes most of the long-term gain: removing it or the centroid update raises errors at 1000 ms, and the model stays competitive with spatio-temporal baselines at that horizon.","The model reaches these results with a small parameter count, which the paper attributes to preserved equivariance."],"supporting_citations":[{"why":"Equivariant multi-agent motion baseline that aggregates only spatial neighboring edges; the paper compares against it and claims the temporal radial field improves long-term accuracy.","marker":"[34]"},{"why":"Equivariant graph mechanics networks provide the rigid-body dynamics formula and equivariant message passing that the paper replaces with a covariance-matrix attention.","marker":"[35]"},{"why":"E(n) equivariant GNNs supply the radial-field message-passing framework and the equivariance argument under orthogonal transformations.","marker":"[33]"},{"why":"Skeleton-parted graph scattering network is the prior body-partition baseline whose upper/lower grouping is compared against the six-group strategy.","marker":"[28]"},{"why":"MSR-GCN establishes the Human3.6M joint selection and 25 Hz protocol, and is a primary multi-scale GCN baseline.","marker":"[24]"},{"why":"Traj-GCN contributes the learnable fully-connected graph approach and the dataset split reused for Human3.6M and CMU-Mocap.","marker":"[26]"},{"why":"Transformer self-attention motivates the covariance-matrix attention inside the equivariant MLP.","marker":"[36]"}],"fun_headline_variants":["Group graph with radial field wins short-term motion prediction","Equivariant group dynamics model predicts motion more accurately","Group graph hits 9 mm error at 80 ms in motion prediction","Group-based physics model beats prior art in short-term motion","Short-term motion prediction: new group graph network improves accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the auxiliary loss term $L_{\\text{aux}}$ enforces human skeleton kinematic constraints by preserving joint lengths, but the formula printed for it only sums L1 norms of predicted bone vectors without referencing any target length, so as written it would shrink bones toward zero instead of constraining them.","fun_headline_variants_meta":{"raw":{"variants":["Group graph with radial field wins short-term motion prediction","Equivariant group dynamics model predicts motion more accurately","Group graph hits 9 mm error at 80 ms in motion prediction","Group-based physics model beats prior art in short-term motion","Short-term motion prediction: new group graph network improves accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000841,"raw_usage":{"total_tokens":3667,"prompt_tokens":949,"completion_tokens":2718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2650}},"tokens_in":565,"tokens_out":2718,"duration_ms":22438,"temperature":1.0,"reasoning_tokens":2650,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:38:52.688816+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code at the paper's repository and run the published $L_{\\text{aux}}$ formula on predicted sequences: if average bone length over the test set decreases monotonically with training while the loss drops, the kinematic-constraint interpretation fails; alternatively, compare predictions trained with $L_{\\text{aux}}$ to ones trained without it and measure whether per-bone variance relative to ground-truth bone lengths actually shrinks.","supporting_citations":[{"cited_title":"Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning,","cited_arxiv_id":null,"evidence_quote":"Equivariant multi-agent motion baseline that aggregates only spatial neighboring edges; the paper compares against it and claims the temporal radial field improves long-term accuracy."},{"cited_title":"Equivariant graph mechanics networks with constraints,","cited_arxiv_id":null,"evidence_quote":"Equivariant graph mechanics networks provide the rigid-body dynamics formula and equivariant message passing that the paper replaces with a covariance-matrix attention."},{"cited_title":"E (n) equivariant graph neural networks,","cited_arxiv_id":null,"evidence_quote":"E(n) equivariant GNNs supply the radial-field message-passing framework and the equivariance argument under orthogonal transformations."},{"cited_title":"Skeleton- parted graph scattering networks for 3d human motion prediction,","cited_arxiv_id":null,"evidence_quote":"Skeleton-parted graph scattering network is the prior body-partition baseline whose upper/lower grouping is compared against the six-group strategy."},{"cited_title":"Msr-gcn: Multi-scale residual graph convolution networks for human motion prediction,","cited_arxiv_id":null,"evidence_quote":"MSR-GCN establishes the Human3.6M joint selection and 25 Hz protocol, and is a primary multi-scale GCN baseline."},{"cited_title":"Learning trajectory depen- dencies for human motion prediction,","cited_arxiv_id":null,"evidence_quote":"Traj-GCN contributes the learnable fully-connected graph approach and the dataset split reused for Human3.6M and CMU-Mocap."}],"review_version":1}