CRISP: Persistent Concept Unlearning via Sparse Autoencoders

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.

Key Contributions

Automated contrastive activation analysis pipeline that identifies SAE features salient to a target concept while discriminating from benign concepts
CRISP: a parameter-efficient persistent unlearning method that suppresses identified SAE features via LoRA fine-tuning, making safety interventions robust to parameter-level bypass
Feature-level analysis demonstrating semantically coherent separation between target and benign concept directions in the SAE feature space