- 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
- 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.
- 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
- 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.
- 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
- 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.
- 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?
- 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?