Paper: "LLMs + Persona-Plug = Personalized LLMs"
📚 https://arxiv.org/abs/2409.11901
🎯 Problem:
- Personalization crucial for language tasks
- Users with same requirements prefer diverse outputs
- Existing approaches:
- Fine-tuning unique LLM per user (expensive)
- Retrieval-based (lacks continuity, sub-optimal)
🔬 Proposed Solution: Persona-Plug
- Constructs user-specific embedding
- Lightweight plug-in user embedder module
- Models all historical contexts
- Attaches embedding to task input
- LLMs capture user habits without parameter tuning
💪 Advantages:
- Better understanding of user preferences
- More personalized outputs
- No LLM parameter tuning required
