LATTE: Forecasting Peer Anchored Preference Trajectories for Personalized LLM Generation

In three linesLATTE is a personalization framework for frozen LLMs that forecasts user preference trajectories by subtracting comparable peer profiles. A lightweight sequence predictor forecasts the next state, injected via a single anchored soft token. On Amazon Reviews 2023, LATTE achieves ROUGE-L=0.259 vs 0.219 for static profiles.

Summary generated by Claude — human-verified