Latent Preference Modeling for Cross-Session Personalized Tool Calling

LLM agents are increasingly expected to call APIs on behalf of users, but real users rarely spell out every argument they want — they just say "book a flight for my trip" and expect the agent to know they always fly economy. We argue this isn't a memory retrieval problem but a memory abstraction problem: the agent has to figure out which past choices reflect reusable preferences and which were just one-off decisions.
We introduce MPT, a benchmark of 265 multi-session dialogues testing three reasoning types — Preference Recall, Induction, and Transfer — and PRefine, a test-time method that maintains latent preferences as evolving hypotheses through a generate–verify–refine loop. PRefine improves tool-calling accuracy across 8 LLMs while using only 1.24% of the tokens required by full-history prompting. The takeaway: robust personalization depends on capturing the reasons behind user choices, not just the choices themselves.