Transformers learn connectivity on low-dimensional grid graphs but fail on high-dimensional grids or graphs with many disconnected components, with larger models showing better generalization on grids.
Do larger language models imply better generalization? a pretraining scaling law for implicit reasoning, 2025
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
years
2025 2verdicts
UNVERDICTED 2roles
background 1polarities
background 1representative citing papers
Deep sequence models develop geometric memory in embeddings that encodes novel global relationships, transforming l-fold composition tasks into 1-step navigation via a natural spectral bias connected to Node2Vec.
citing papers explorer
-
Transformers Can Learn Connectivity in Some Graphs but Not Others
Transformers learn connectivity on low-dimensional grid graphs but fail on high-dimensional grids or graphs with many disconnected components, with larger models showing better generalization on grids.
-
Deep sequence models tend to memorize geometrically; it is unclear why
Deep sequence models develop geometric memory in embeddings that encodes novel global relationships, transforming l-fold composition tasks into 1-step navigation via a natural spectral bias connected to Node2Vec.