Capacity scaling and generalization tests indicate that transformers without chain of thought learn latent two-hop questions by storing each fact twice (two-function composition), while chain-of-thought models store facts once.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Examining Two Hop Reasoning Through Information Content Scaling
Capacity scaling and generalization tests indicate that transformers without chain of thought learn latent two-hop questions by storing each fact twice (two-function composition), while chain-of-thought models store facts once.