1. From Edge Facts to a Latent Map
The paper’s claim is easiest to see as a conversion pipeline: local supervision on graph edges first writes memories into weights, then those memories reorganize into a geometry that can answer queries not seen directly.
2. Lookup Chain vs Geometric Shortcut
Toggle the retrieval story. In the associative view, answering a path query means traversing stored links. In the geometric view, the same answer emerges from a much shorter move in representation space.
3. Heatmaps of Memory Structure
These matrices use realistic mock values to contrast raw edge co-occurrence, a narrow lookup-style memory, and a smoother latent similarity map that reflects broader graph structure.
4. What Improves, and What Still Feels Fragile
The mock result panels below reflect the qualitative story from the discussion: held-out path queries improve when geometry emerges, but synthetic graph wins do not automatically settle the large-model reasoning debate.