Language models' next-token predictions for related knowledge are connected by near-linear transformations that persist through fine-tuning, explaining both compositional generalization and hallucination.
Evaluating the ripple effects of knowledge editing in language models
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Linear Correlation in LM's Compositional Generalization and Hallucination
Language models' next-token predictions for related knowledge are connected by near-linear transformations that persist through fine-tuning, explaining both compositional generalization and hallucination.