RePAIR: Interactive Machine Unlearning through Prompt-Aware Model Repair

Researchers introduce RePAIR, a framework enabling users to instruct large language models to forget harmful knowledge, misinformation, and personal data through natural language prompts at inference time. The system uses a training-free method called STAMP that manipulates model activations to achieve selective unlearning with minimal computational overhead, outperforming existing approaches while preserving model utility.

RePAIR addresses a critical gap in AI safety and user privacy by decentralizing machine unlearning from model providers to end users. Traditional unlearning approaches require complete model retraining, curated datasets, and provider intervention—barriers that exclude most users from controlling their own data. This research shifts the paradigm toward interactive, on-demand unlearning at inference time, fundamentally changing how users can manage AI systems' knowledge retention.

The framework's technical innovation lies in STAMP, a training-free method that redirects neural network activations toward a refusal subspace using pseudoinverse mathematics. By avoiding full retraining, STAMP reduces computational complexity from cubic to near-linear scaling, enabling efficient deployment on consumer devices. This efficiency represents a significant engineering achievement that makes practical, user-driven model editing feasible at scale.

The implications extend across multiple domains. For AI safety, on-device unlearning reduces risks from harmful knowledge absorption while respecting privacy regulations like GDPR. For developers, this enables personalized model behavior without redeploying infrastructure. For enterprises, it provides mechanisms to correct misinformation or remove inadvertently trained sensitive data post-deployment—a capability currently unavailable in production systems.

Market impact remains indirect but meaningful. The framework doesn't directly monetize but enhances AI system reliability, potentially increasing user trust and adoption. The methodology could accelerate development of trustworthy AI systems, influencing how next-generation foundation models are designed. Success at scale would validate that selective unlearning is practical, potentially reshaping how AI companies approach safety and compliance.

• →RePAIR enables end users to instruct LLMs to forget harmful knowledge through natural language prompts without model retraining.
• →STAMP's training-free approach reduces computational complexity by ~3x compared to training-based baselines, enabling on-device unlearning.
• →Framework achieves near-zero forget scores while preserving model utility, validating practical selective unlearning at scale.
• →Decentralizes machine unlearning from model providers to users, advancing user control over AI systems and privacy compliance.
• →Low-rank optimization technique enables efficient implementation on consumer hardware, broadening accessibility of model editing capabilities.