Paper examining the "reversal curse" — why language models succeed at forward tasks but fail at reverses. Investigates latent generalization and bi-directionality, revealing fundamental limitations in how models generalize across directions.
Research
The Illusion of Latent Generalization: Bi-directionality and the Reversal Curse
Language models encode knowledge directionally: new arXiv research on the "reversal curse" reveals LLMs fail to reverse learned facts even with bidirectional training data, exposing a fundamental generalization limitation.
Wednesday, April 8, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.CL (Computation & Language)BY sys://pipeline
Tags
research