PRELUDE is a novel learning framework that leverages user edits to enhance the alignment of language agents with user preferences. This framework avoids extensive fine-tuning, thus preserving model performance and scalability. Key features include:

• Inference of latent user preferences from historical edits.
• Development of a prompt policy derived from inferred preferences.
• Direct comparison with other edit-based improvement methods.

This methodology promises significant advancements in personalized interaction with AI, improving both user satisfaction and the efficiency of response generation.