Title: Don't Forget Your Embeddings: Robust Knowledge Erasure via Precise Editing of Embeddings

Abstract:

As language models become more deeply integrated into practical applications, the capacity to permanently remove specific information from them has emerged as a vital requirement for ensuring safety and regulatory compliance. While current prominent approaches aim for lasting deletion by modifying model parameters, this target knowledge frequently resurfaces via adversarial prompting or through relearning processes. We posit that this vulnerability partly arises because prevailing techniques neglect the embedding layer. To overcome this gap, we present EMBER (EMBedding ERasure), a modular, plug-and-play solution that employs Sparse Matrix Factorization to accurately eliminate concept-related features from token embeddings.

Our extensive testing across a wide array of concepts on the Gemma-2-2B-it and Llama-3.1-8B-Instruct models reveals that integrating EMBER with existing erasure methods consistently enhances both the effectiveness and precision of knowledge removal across various task formats, while incurring negligible loss of coherence. Furthermore, EMBER significantly bolsters resistance to relearning; it slashes regained accuracy by as much as 50%, capping the recovery rate at 35% for Llama, a stark contrast to the 70%-76% recovery seen with earlier methods. Additional analysis indicates that any coherence penalty is highly localized, impacting only a narrow selection of tokens unique to the targeted concepts. This study confirms that precise, embedding-level intervention is essential for durable concept erasure and highlights how such augmentation can strengthen current methodologies.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC