acceptodds
Withdrawn from ICLR 2027

EqDynHyper Equivariant Dynamic Hypergraph for Group-Level Motion Prediction and Group Discovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-agent motion forecasting must respect geometric symmetries while modeling both pairwise interactions and emergent collective group dynamics. Existing Euclidean-equivariant predictors rely on pairwise message passing and have no native group-level primitive, whereas hypergraph forecasters that do model groups are not Euclidean-equivariant. We present EqDynHyper, an equivariant hypergraph network with two disentangled outputs: trajectory forecasts, whose accuracy is insensitive to grouping quality and whose efficiency comes from equivariant centroid aggregation over hyperedges, and agent-group partitions, recovered without membership supervision through Euclidean-invariant soft hyperedge assignments. Building on the scalar-vector separation of GVP-GNN and PaiNN, we add an explicit affine Point type and the closure rules needed for higher-order aggregation, so that trajectories are equivariant and assignments invariant by construction. Assignments are computed once by default and can optionally be re-inferred at every layer. We find that competitive (softmax) normalization of the assignment matrix, rather than auxiliary entropy losses, is what prevents hyperedge collapse. EqDynHyper replaces vector-valued pairwise message passing with hyperedge aggregation; a scalar pairwise topology term remains , and at inference is faster with less memory than EqMotion. On synthetic collective dynamics EqDynHyper lowers ADE by relative to EqMotion, and on ETH-UCY by 5.0 percent under a unified retraining protocol; on bond-dominated MD17 molecules it is less accurate than pairwise models, which we report together with other failure modes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Trajectory Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LUS3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EqDynHyper combines Euclidean-invariant soft agent-to-hyperedge assignments with equivariant centroid-based message passing, using point/vector/scalar operations. Assignments are static by default and optionally recomputed across layers. Forecasting and membership-unsupervised grouping are explicitly separate outputs. Experiments report substantial synthetic ADE improvements over EqMotion, a roughly 5% ETH–UCY ADE improvement under modified, jointly retrained preprocessing, and constant-factor inference savings. The paper also establishes that accurate grouping is unnecessary for these forecasting gains and reports inferior performance on three MD17 molecules.

Critical strengths
  1. The point/vector/scalar construction cleanly couples invariant assignments to equivariant forecasts. Appendix A checks the affine centroid normalization and padding requirements; the reported untrained-network rotation, reflection, translation, and permutation tests meaningfully complement the proof.
  2. There are substantial measured benefits and unusually candid controls: synthetic ADE improves from 0.0212 to 0.0072 at N=12, and matched-parameter inference is reported 9.4× faster at N=400. K=1/random controls, MD17 failures, constant-factor rather than asymptotic scaling, and limited dynamic-topology gains are all disclosed. Unsupervised grouping also reaches 0.859±0.047 F1 without load regularization, over five seeds.
Critical weaknesses
  1. The forecasting improvement is not isolated to the proposed interaction mechanism. K=1/random assignments nearly match the full model, while no otherwise-matched model without inter-agent aggregation is evaluated (Table 2; §5.5). The authors explicitly acknowledge this. Nevertheless, since temporal lifting, velocity residuals, and decoding also differ from EqMotion, the large synthetic improvement does not yet establish what centroid aggregation contributes beyond an efficient implementation.
  2. The independent group-discovery contribution lacks a convincing matched comparison with dedicated latent-hypergraph methods. DynGroupNet is deferred, and SPHINX (ICML 2025), which studies unsupervised hypergraph recovery from rotating-particle trajectories, is absent. The acknowledged unreliable GroupNet reproduction further limits that comparison. The softmax anti-collapse interpretation rests on one sigmoid run and grouping F1 rather than direct collapse statistics. Reproduction is additionally hindered by missing synthetic-generator/split specifications and exact custom ETH–UCY preprocessing details; code is promised only upon acceptance.
Questions for authors
  1. What remains of the forecasting gain when inter-agent aggregation is removed while temporal lifting, velocity residuals, decoder, parameter budget, and training protocol are matched? This would distinguish the proposed aggregation's contribution from the remaining architecture.
  2. At the same supplied group count, how does group recovery compare with a dedicated method such as SPHINX or DynGroupNet under matched data and metrics? Can multi-seed occupancy/collapse statistics substantiate the softmax interpretation separately from F1?
  3. Can you supply the Group-Hard generator equations, sample counts, horizons/timestep and splits; the ETH–UCY neighbor-selection and validation-splitting rules; and the GPU, precision, and timing procedure for Table 6?
Reviewer_JR8KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EqDynHyper for multi-agent motion forecasting and group discovery. The method infers soft agent-to-hyperedge assignments from Euclidean-invariant features and uses Point, Vector, and Scalar type constraints to construct equivariant hypergraph message passing and trajectory decoding. It replaces pairwise interactions in the vector branch with hyperedge aggregation to reduce computational costs while producing coordinate-independent group assignments. The authors evaluate forecasting accuracy, grouping quality, and computational efficiency on synthetic collective dynamics, ETH–UCY, and MD17, with ablations examining grouping and dynamic topology.

Critical strengths
  1. The geometric construction is clear and verifiable. Explicitly distinguishing Points, Vectors, and Scalars makes the transformation properties of hyperedge centroids and subsequent updates transparent. The appendix provides a layer-wise equivariance proof and numerical checks covering padded variable-size batches, connecting the theoretical guarantees with implementation details.

  2. The method demonstrates practical computational benefits and candidly discusses its limitations. Hyperedge aggregation in the vector branch reduces latency and memory usage in large-scale forward-pass benchmarks. The authors also report the remaining quadratic complexity, weaker performance on MD17, and the limited dependence of forecasting accuracy on grouping quality, providing useful evidence about the method’s applicability.

Critical weaknesses
  1. Single-hyperedge and random-grouping baselines perform similarly to the full model. Without a matched centroid-pooling baseline, the forecasting benefit of learned hypergraph structure beyond global equivariant aggregation remains unclear.

  2. Although the method demonstrates practical speedups, large-scale benchmarks report efficiency without prediction accuracy. The accuracy–efficiency trade-off at larger agent counts remains unverified.

  3. Group discovery requires a predefined KK. The reported sensitivity analysis shows substantial degradation when KK differs from the true group count, but its dependence on the regularization setting remains unclear.

Questions for authors
  1. Could a matched centroid-pooling baseline clarify whether learned hypergraph structure improves forecasting beyond global equivariant aggregation?

  2. Could the authors evaluate prediction accuracy alongside runtime and memory at larger agent counts to establish the accuracy–efficiency trade-off?

  3. When the true group count is unknown, how is KK selected? Does grouping remain sensitive to misspecified KK under the recommended regularization setting?

Reviewer_2PPsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Large-scale multi-agent motion prediction faces many challenges, such as the high cost of pairwise equivariant messages. This paper proposes the EqDynHyper framework. The framework reconstructs the pairwise messages of the vector branch into hyperedge aggregation and anchors equivariant centroids through Euclidean-invariant soft group assignments, thereby decoupling pairwise topology computation from group-level vector aggregation. Experiments verify the effectiveness of the framework in equivariant outputs, group discovery, and inference efficiency.

Critical strengths
  1. The paper's classification and positioning of related work are well-organized, helping readers quickly grasp the differences and entry points of this paper relative to existing methods.
  2. The proposed method is built on a clear geometric type system and equivariance proofs, with theoretical derivations and numerical validations supporting each other, making the argumentation relatively rigorous.
  3. The experiments cover multiple types of benchmarks, resulting in a relatively comprehensive evaluation.
Critical weaknesses
  1. The paper uses equivariance as motivation but mainly relies on literature conventions and a self-constructed variant on a single synthetic task. Meanwhile, the authors also admit in Section 5.5 that equivariance itself does not constitute an advantage over equivariant baselines.
  2. The introduction is scattered and resembles related work. The listing of shortcomings lacks convergence, and details such as the type system and pooling analogies repeat later sections, weakening the persuasiveness of the method design.
  3. The authors acknowledge in the paper that the overall complexity remains O(N^2) and that the speedup is a constant factor. Meanwhile, the accuracy of K=1 and random grouping is close to that of the full model, weakening the necessity of the hypergraph structure.
  4. The paper's group assignment is formally close to soft clustering/soft k-means rather than discovering group structure.
  5. In the main experiments, K is basically set to the ground-truth number of groups, and F1 drops noticeably when K deviates. The sensitivity experiments are again conducted only on a small-scale synthetic set, with non-recommended hyperparameters, and as single runs, weakening the claim of unsupervised group discovery.
  6. Only forward latency and memory are reported, while training time/throughput are missing, and the measurement conditions (hardware, batch, precision) and baseline comparisons are also insufficient.
Questions for authors
  1. Does this paper claim that "equivariance is a task necessity," or that "equivariance is a design constraint, and the real gain comes from hyperedge aggregation"?
  2. How should the core gap of this paper relative to existing methods be summarized?
  3. How much of the prediction accuracy gain can be attributed to hyperedge aggregation itself?
  4. What is the substantive difference between the assignment in this paper and "soft clustering with known K followed by re-aggregation," and has such a comparison been conducted?
  5. Does "unsupervised group discovery" hold only under the premise of a given ground-truth K?
  6. What are the measurement conditions of Table 6, and is the training cost consistent with the inference benefit?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.