Informational Antilocality and the Locality Bias in LLMs
Computation and Language
Summary
The authors studied how well transformer-based language models can learn special languages called k-antilocal languages, where there is no connection between any sequence of k symbols. They created these languages with different values of k and found that while the models eventually learn them equally well, it takes longer to learn languages with higher k. This suggests that learning relationships between far-apart parts in language is harder, but the models still succeed in the end.
Authors
Andrew McInnerney, Shane Storks, Steven Abney, Richard L. Lewis
Abstract
We consider the ability of transformer-based language models (LLMs) to learn what we call k-antilocal languages, i.e., languages that have no mutual information across any span of $k$ contiguous symbols. We construct such languages with increasing $k$, finding that LLMs trained on them achieve comparable cross-entropy loss regardless of antilocality, but converge more slowly on more antilocal languages. Our findings support the idea that non-local dependencies are more difficult to learn, but the evidence for this bias comes from learning speed rather than learning success.