CRISP: Persistent Concept Unlearning via Sparse Autoencoders

TL;DR

CRISP is a novel, parameter-efficient method that uses sparse autoencoders to achieve persistent, precise unlearning of harmful concepts in large language models, enhancing safety without sacrificing utility.

Contribution

CRISP introduces a new approach for persistent concept unlearning in LLMs by automatically identifying and suppressing salient features across layers, outperforming prior methods.

Findings

CRISP successfully removes harmful knowledge in LLMs from the WMDP benchmark.

CRISP preserves general and in-domain capabilities after unlearning.

Feature analysis shows semantically coherent separation of concepts.

Abstract

As large language models (LLMs) are increasingly deployed in real-world applications, the need to selectively remove unwanted knowledge while preserving model utility has become paramount. Recent work has explored sparse autoencoders (SAEs) to perform precise interventions on monosemantic features. However, most SAE-based methods operate at inference time, which does not create persistent changes in the model's parameters. Such interventions can be bypassed or reversed by malicious actors with parameter access. We introduce CRISP, a parameter-efficient method for persistent concept unlearning using SAEs. CRISP automatically identifies salient SAE features across multiple layers and suppresses their activations. We experiment with two LLMs and show that our method outperforms prior approaches on safety-critical unlearning tasks from the WMDP benchmark, successfully removing harmful knowledge while preserving general and in-domain capabilities. Feature-level analysis reveals that CRISP achieves semantically coherent separation between target and benign concepts, allowing precise suppression of the target features.