View PDF HTML (experimental)

Abstract:We study factual edit propagation in a controlled synthetic knowledge-graph QA setting, comparing four architectures that cross loop recurrence with shared memory access: dense (Dense), looped (Loop), dense with shared memory (Dense+Mem), and looped with shared memory (LMC). Dense fits in-distribution 2-hop compositions but fails OOD; both looped recomputation and memory rereading cross this OOD grokking barrier, indicating that repeated shared access -- not a specific architecture -- is the common ingredient for learning. Editing, however, separates the substrates along a different axis. On a shared pre-edit-correct ID set, a single-row factual edit propagates strongly in the two memory-bearing cells (LMC 0.78-0.92, Dense+Mem 0.71-0.96) and only weakly in the others (Loop 0.04-0.30, Dense 0.00-0.03); the separation is statistically clean (Mann-Whitney p=0.008 between memory and non-memory cells, p=0.55 between the two memory cells, though n=5 vs n=5 is underpowered to rule out a moderate gap). In LMC, atomic facts localize to dominant memory sites that composition rereads, and a one-row value edit on LMC's own pre-edit-correct probes achieves 100% direct success with mean 0.989 intended propagation while moving unrelated facts ~0.1% of the time (matched specificity across substrates pending). A coarse hold-answer-subspace (HOLDANS) interchange diagnostic is consistent with this ordering, suggesting that what separates the substrates is when an edited fact is injected and how much computation remains afterwards to reuse it. These results dissociate learning competence from editing affordance: repeated shared access suffices to grok, but edit propagation depends on whether the substrate exposes a fine-grained addressable memory the forward computation can write to and later reread -- an affordance that loop recurrence provides only partially.

Submission history

From: Yanan Niu [view email]
[v1] Wed, 17 Jun 2026 15:52:08 UTC (30 KB)