Research map / Large language models / Language, safety and interpretability
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
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| migrate soft prompts with PUMA adapter instead of retraining from scratch | personalized soft prompts · Amazon (1B->3B migration) | RMSE: 0.9414 -> 0.9135 | Improved |
| migrate soft prompts with PUMA adapter instead of retraining | personalized soft prompts · MIND (1B->3B migration) | uAUC: 0.5289 -> 0.6552 | Improved |
| migrate with PUMA instead of retraining | personalized soft prompts · Amazon | training time (h): 24.0 -> 0.48 | Improved |
| use K-means + variance stratification instead of random sampling at same budget (ablation) | PUMA user selection · Amazon (2k-user selection) | RMSE: 0.9419 -> 0.9315 | Improved |
| use K-means + variance stratification instead of K-means stratified alone (ablation) | PUMA user selection · Amazon (2k-user selection) | RMSE: 0.9546 -> 0.9315 | Improved |
| use K-means + variance stratification instead of random on Yelp (ablation) | PUMA user selection · Yelp (2k-user selection) | RMSE: 1.1146 -> 1.1111 | No clear effect |
| use K-means + variance stratification vs Loss Bucketing (ablation) | PUMA user selection · MIND (1.5k-user selection) | uAUC: 0.5845 -> 0.6344 | Improved |
| aggregate prompts from two source models (Llama + StableLM) | PUMA migration to Phi-3 · Amazon aggregated migration to Phi-3 | RMSE: 0.9293 -> 0.9217 | Improved |
| chain migration across five models | PUMA · Amazon chain migration (Llama start) | RMSE: 0.9348 -> 0.9277 | Improved |
| migrate from weaker source model to stronger target | PUMA · Amazon cross-architecture (Figure 4) | RMSE gain | Mixed |
Closest papers by meaning
- Rethinking Personalization in Large Language Models at the Token Level (ICML 2026 · similarity 0.87)
- Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass (ICLR 2025 · similarity 0.86)
- Efficient and Privacy-Preserving Soft Prompt Transfer for LLMs (ICML 2025 · similarity 0.86)
- System Prompt Optimization with Meta-Learning (NeurIPS 2025 · similarity 0.86)
- PortLLM: Personalizing Evolving Large Language Models with Training-Free and Portable Model Patches (ICLR 2025 · similarity 0.86)
- HYDRA: Model Factorization Framework for Black-Box LLM Personalization (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
