Generalization across graph topologies and scales

Ascertain whether the observed implicit in-weights reasoning and geometric memory in Transformer and Mamba models extend beyond path-star and tree-star graphs to other graph topologies and to graphs of different sizes.

Background

The positive results in the paper are demonstrated primarily on path-star (and additionally tree-star) graphs, showing strong in-weights generalization on held-out leaves and large graphs. Whether this behavior persists on broader classes of graphs or at different scales is not yet established.

Clarifying this would define the scope of applicability of geometric memory in parametric models and inform the design of benchmarks and training protocols.

References

It is unclear how well this generalizes to other topologies, and to graphs of other sizes.

— Deep sequence models tend to memorize geometrically; it is unclear why  (2510.26745 - Noroozizadeh et al., 30 Oct 2025) in Section: Limitations (bullet 1)

Given the partial nature of these tests, however, generalisation across graphs remains an open question in our work.

— A Code-Agnostic Graph Neural Network Decoder from the Detection Error Model  (2610.01683 - Astolfi et al., 1 Oct 2026) in Section 1, Introduction, paragraph beginning “There are two main advances in this work.”

Taken together, these results demonstrate substantial variation across models and graph settings, while leaving the sources of their errors unresolved.

— MVGraph: Benchmarking LVLMs on Evidence Composition across Multi-View Visual Graph  (2610.12961 - Zhu et al., 9 Oct 2026) in Section 5.1, “End-to-End Graph-Theoretic Performance”