대규모 언어모델에서 메모라이제이션 기반 언러닝 = Measuring Memorization-based Unlearning in Large Language Models

다국어 초록 (Multilingual Abstract)

Large Language Models (LLMs) exhibit remarkable generative capabilities; however, they also demonstrate memorization behaviors, which can lead to serious ethical and legal concerns, such as the leakage of private information and the reproduction of copyrighted content. To address these issues, machine unlearning has been actively studied as a mitigation strategy. Nevertheless, unlearning becomes increasingly challenging as the degree of memorization grows. In particular, strongly memorized samples are difficult to remove effectively, while overly aggressive unlearning often results in significant degradation of model utility.

In this work, we propose a framework that measures memorization at the sample level and leverages this information to guide differentiated unlearning strategies. However, accurately measuring memorization in large language models remains challenging, as existing approaches—such as leave-one-out retraining or influence function–based methods—are often computationally expensive and numerically unstable at scale. To address these limitations, we introduce Paraphrase Gradient Alignment (PGA), a novel gradient-based memorization metric. PGA quantifies the degree of gradient alignment induced by an original sentence and its near-duplicate paraphrases, providing a proxy for the sample’s memorization strength. Importantly, PGA relies solely on gradients computed from the current model parameters, enabling efficient memorization estimation without retraining, Hessian approximation, or tracking the training trajectory.

Experiments on the TOFU and MUSE benchmarks demonstrate that PGA achieves strong discriminative performance in Membership Inference Attack (MIA) tasks, outperforming or matching existing memorization metrics. Building on this measurement, we design curriculum-based and hybrid unlearning strategies that adapt unlearning behavior according to memorization strength. Our results show that parameter-update-based unlearning is effective for weakly memorized samples, while strongly memorized samples benefit from combining test-time unlearning to suppress direct output, thereby preserving model utility.

Overall, this work reframes unlearning as a memorization-aware, data-adaptive problem rather than a one-size-fits-all algorithmic solution. By tightly coupling memorization measurement with strategic unlearning, we provide a practical and principled framework for effective unlearning in large language models.

목차 (Table of Contents)

• 1서론 11
• 2관련 연구 15
• 2.1 메모라이제이션 15
• 2.2 거대 언어모델 언러닝 17
• 2.3 Approximate Unlearning 19
• 3메모라이제이션 측정과 언러닝 23
• 3.1 메모라이제이션 측정을 위한 문장 생성 24
• 3.2 Gradient Alignment 기반 메모라이제이션 측정 26
• 3.3 하이브리드 언러닝 27
• 실험 30
• 4.1 실험 환경 30
• 4.1.1 데이터셋 31
• 4.1.2 평가 메트릭 32
• 4.1.3 구현 세부사항 34
• 4.2 유사 중복 문장 생성 36
• 4.3 Membership Inference 실험 37
• 4.4 하이브리드 언러닝 39
• 5 결론 및 향후 연구 47