Effect of stronger structural-inference backbones on SIHA refinement

Determine whether pairing the SIHA structure-guided hidden-state predictor (HSP-sg) and its iterative state–structure refinement mechanism with a stronger structural-inference backbone improves reconstruction, forecasting, and structural-inference performance beyond the current Neural Relational Inference backbone.

Background

SIHA uses Neural Relational Inference (NRI) as both its structural-inference and future-prediction backbone. The authors observe that the benefit of structure-guided refinement varies across the Springs, Charged Particles, and Kuramoto systems, potentially because NRI infers structures with different levels of accuracy on these datasets.

The paper conjectures that a stronger structural-inference backbone could improve HSP-sg, but does not test this possibility. The unresolved issue is therefore whether the SIHA refinement mechanism is more effective when coupled with a structural-inference model that provides more accurate interaction estimates.

References

We conjecture that this difference may be related to the quality of the structures inferred by the underlying NRI model: its structural inference accuracy is lower on Charged and Kuramoto than on Springs, which may limit the benefit of the NRI-guided iterative refinement in HSP-sg. Since the SIHA refinement mechanism only requires a structural estimate together with future-state prediction, it could in principle be paired with other compatible structural-inference backbones. We therefore expect that a stronger backbone may further improve the effectiveness of HSP-sg, although this remains to be verified experimentally.

— Structural Inference under Hidden Agents  (2609.18045 - Gong et al., 16 Sep 2026) in Section 4.1, subsection “Overall Performance”