Erasing Conceptual Knowledge from Language Models

TL;DR

This paper presents ELM, a novel method for concept-level unlearning in language models that reduces undesired concept generation while preserving overall performance and robustness.

Contribution

Introduces ELM, a principled approach leveraging the model's own introspective capabilities for targeted concept unlearning via low-rank updates.

Findings

ELM effectively erases targeted concepts from language models.

Models with ELM show near-random performance on erased concepts.

ELM preserves model performance on unrelated tasks and enhances robustness.

Abstract

In this work, we introduce Erasure of Language Memory (ELM), a principled approach to concept-level unlearning that operates by matching distributions defined by the model's own introspective classification capabilities. Our key insight is that effective unlearning should leverage the model's ability to evaluate its own knowledge, using the language model itself as a classifier to identify and reduce the likelihood of generating content related to undesired concepts. ELM applies this framework to create targeted low-rank updates that reduce generation probabilities for concept-specific content while preserving the model's broader capabilities. We demonstrate ELM's efficacy on biosecurity, cybersecurity, and literary domain erasure tasks. Comparative evaluation reveals that ELM-modified models achieve near-random performance on assessments targeting erased concepts, while simultaneously preserving generation coherence, maintaining benchmark performance on unrelated tasks, and exhibiting strong robustness to adversarial attacks. Our code, data, and trained models are available at https://elm.baulab.info