[arXiv]score: 0.44
LATTE Models User Preference as Peer-Relative Latent Trajectory for LLM Personalization
May 27, 2026
LATTE separates stable identity from recent preference drift by computing per-session states as deviations from time-masked peer baselines on the same items, then forecasting the next state with a lightweight sequence predictor to condition frozen LLMs.
cs.CL
HOW THIS AFFECTS YOU
●
builderWorth evaluating if you're building personalized generation on top of frozen LLMs and current static profile approaches aren't capturing recency effects.
●
researcherThe peer-anchored relative state formulation is a novel decomposition that disentangles item content effects from user-specific drift in personalization representations.