Tag
The paper introduces MWE-ECL, a diagnostic tool to test whether AI models can use long-range context to override local semantic priors in multiword expressions, highlighting gaps between recoverability and behavioral influence.