Predicting Linear Relational Decodability via the Entity Lens
October 21, 2026
The entity lens method predicts how well language models map subject-to-object relations using only entity-type information. By training on neutral prompts like '{} is known for', this readout estimates the faithfulness of linear relational embeddings across 28 different relations without needing relation-specific data.
HOW THIS AFFECTS YOU
●
researcherYou can use this to predict the reliability of relational embeddings in your models without explicit relation labels.