ScaleToT Generalizes LLM Reasoning for Billion-Scale User Modeling

Accurate user modeling typically relies on extensive interaction histories, which are often unavailable for a vast number of low-activity users. While Large Language Models (LLMs) can infer user states from static profiles, this approach becomes unreliable with sparse data and prohibitively expensive at a billion-user scale. ScaleToT addresses this by learning structured reasoning from a small, LLM-processed user subset and then generalizing it to the entire low-activity user base. ScaleToT enhances reasoning reliability by constructing typed user-state chains using an entropy-guided Tree-of-Thought (ToT) refinement procedure. This teacher-curated reasoning is then used to train a student model on static profiles through supervised fine-tuning (SFT) and a novel Outcome-Driven Segment-Aware Implicit Reward Policy Optimization (OSIPO). This two-stage process allows the student model to acquire the complex reasoning capabilities from sparse data. Finally, ScaleToT transfers the student model's reasoning representations to a lightweight profile encoder. This encoder provides shared reasoning signals for the remaining users without requiring direct LLM inference, drastically cutting down computational costs. Evaluated in a billion-scale advertising deployment for lifetime value (LTV) prediction, ScaleToT achieved a 6.738% increase in LT30, demonstrating its effectiveness and efficiency in modeling low-activity users.