Publications
ACL 2026 FindingsThe Model Agreed, But Didn’t Learn: Diagnosing Surface Compliance in Large Language Models
We propose SA-MCQ, a diagnostic framework that reveals surface compliance in knowledge editing, where models appear to accept edits without truly internalizing them.
TMLR 2026UltraEdit: Training-, Subject-, and Memory-Free Lifelong Editing in Language Models
UltraEdit is a lifelong editing framework for language models that removes the need for extra training, subject-specific memory, and memory storage.
