Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Ziyi Zhao, Chongming Gao, Yang Zhang, Haoyan Liu, Weinan Gan, Huifeng Guo, Yong Liu, Fuli Feng

TL;DR

This work tackles the problem of migrating thousands of user-specific soft prompts across incompatible foundation models without full retraining. It introduces PUMA, a lightweight prompt migration adapter combined with a group-based user selection strategy to efficiently bridge semantic gaps between source and target models. Empirical results across three large-scale datasets show that PUMA matches or exceeds retraining performance while reducing computational cost by up to 98%, and it generalizes across diverse architectures and advanced migration topologies like chained and aggregated migrations. By decoupling personalized assets from specific models, PUMA offers a practical, scalable path for sustainable personalization in evolving LLM ecosystems.

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.

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

TL;DR

This work tackles the problem of migrating thousands of user-specific soft prompts across incompatible foundation models without full retraining. It introduces PUMA, a lightweight prompt migration adapter combined with a group-based user selection strategy to efficiently bridge semantic gaps between source and target models. Empirical results across three large-scale datasets show that PUMA matches or exceeds retraining performance while reducing computational cost by up to 98%, and it generalizes across diverse architectures and advanced migration topologies like chained and aggregated migrations. By decoupling personalized assets from specific models, PUMA offers a practical, scalable path for sustainable personalization in evolving LLM ecosystems.

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.
Paper Structure (29 sections, 3 equations, 6 figures, 4 tables)

This paper contains 29 sections, 3 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Personalization in Large Language Models
  4. Coreset Selection
  5. Knowledge Transfer in Soft Prompt
  6. Methodology
  7. Problem Formulation
  8. Personalization with Soft Prompts
  9. Cross-Model Prompt Migration
  10. Adapter Framework
  11. Extension to Advanced Migration Topologies.
  12. Efficient Migration via Group-based User Selection
  13. Stage 1: Cluster for diversity.
  14. Stage 2: Grouped sampling by variance.
  15. Experiments
  16. ...and 14 more sections

Figures (6)

  • Figure 1: The "1+N" system (left) and an illustration of the migration of personalized prompts to a new LLM (right).
  • Figure 2: Illustration of PUMA, consisting of a group-based user selection strategy and a migration adapter. Users are first clustered via $K$-means on personalized embeddings, then sub-grouped by output variance, from which training users are sampled.
  • Figure 3: Performance of random user sampling on Amazon.
  • Figure 4: Performance gain heatmap for different architectures. (Gain = full retrained RMSE / migrated RMSE)
  • Figure 5: Stability in chain migration (started from Llama).
  • ...and 1 more figures