[논문 리뷰] Don't Start Over: A Cost-Effective Framework for Migrating Personalized Prompts Between LLMs

Ziyi Zhao, Chongming Gao외 6명|arXiv (Cornell University)|2026. 01. 17.

Topic Modeling인용 수 0

한 줄 요약

PUMA은 서로 호환되지 않는 LLM 간의 사용자 특화 소프 프롬프트를 가볍게 만든 어댑터와 그룹 기반 사용자 선택을 사용하여 이식하며, 전체 재훈련에 필적하거나 이를 능가하는 성능을 유지하면서 비용을 최대 98%까지 절감합니다.

ABSTRACT

Personalization in Large Language Models (LLMs) often relies on user-specific soft prompts. However, these prompts become obsolete when the foundation model is upgraded, necessitating costly, full-scale retraining. To overcome this limitation, we propose the Prompt-level User Migration Adapter (PUMA), a lightweight framework to efficiently migrate personalized prompts across incompatible models. PUMA utilizes a parameter-efficient adapter to bridge the semantic gap, combined with a group-based user selection strategy to significantly reduce training costs. Experiments on three large-scale datasets show our method matches or even surpasses the performance of retraining from scratch, reducing computational cost by up to 98%. The framework demonstrates strong generalization across diverse model architectures and robustness in advanced scenarios like chained and aggregated migrations, offering a practical path for the sustainable evolution of personalized AI by decoupling user assets from the underlying models.

연구 동기 및 목표

• Address the vulnerability of soft prompts to foundation model upgrades in 1+N personalization systems.
• Develop a cost-efficient method to migrate thousands of user prompts across different models without full retraining.
• Propose a parameter-efficient adapter and a group-based user selection strategy to ensure semantic alignment and scalability.
• Demonstrate robustness across diverse model architectures and advanced migration topologies (chained and aggregated).
• Show that migration can outperform retraining from scratch while decoupling user assets from models.

제안 방법

• Introduce an adapter-based migration function Phi parametrized by theta that maps source prompts to target prompts across models with different embedding dimensions.
• Train Phi end-to-end on a target-model task loss while keeping both the source prompts and models frozen.
• Use a lightweight feed-forward adapter with residual connections and Layer Normalization as Phi.
• Implement a two-stage group-based user selection: K-means clustering for diversity, followed by variance-based grouped sampling to form a small representative training subset.
• Extend PUMA to advanced migration topologies: chained migration and aggregated migration by concatenating prompts from multiple sources.
• Evaluate on three large-scale datasets (Amazon, MIND, Yelp) across multiple model families to show generalization and robustness.

실험 결과

연구 질문

• RQ1 RQ1: How effective is PUMA at migrating user personalization compared to full retraining?
• RQ2 RQ2: How efficiently does the group-based user selection reduce costs while preserving performance?
• RQ3 RQ3: Does PUMA generalize across diverse model architectures and families?
• RQ4 RQ4: How robust is PUMA in advanced migrations like chained and aggregated migrations?

주요 결과

• PUMA matches or surpasses full retraining while significantly reducing computation (up to 98% cost reduction).
• Efficiency gains are achieved via the group-based user selection strategy, outperforming random sampling within the same budget.
• PUMA generalizes well across diverse architectures and model families, with migrations often yielding gains over retraining.
• In chained migrations, performance remains robust across multiple successive model updates; aggregated migrations from multiple sources improve personalization by fusing knowledge.
• The aggregated-source setting can surpass single-source migrations, demonstrating knowledge synergy from multiple foundation models.