AI assistants often fail to match how a user wants help: sometimes a short answer is needed, other times a deep explanation. Existing methods can specify desired behaviors but cannot tell which parts of a response to keep or change.
The authors introduce Persona Learning from Feedback (PLF). It trains independent lightweight adapters (LoRA) for different behavior styles. Each adapter generates its own responses and receives feedback focused on responses needing improvement. At inference, a router chooses or combines adapters based on the request.
PLF improves how well the assistant's behavior fits user preferences compared to strong baselines. The learned adapters can be reused across users and combined for multiple preferences, without requiring per-user fine-tuning.