🎉 Our paper “TimeSRL: Generalizable Time-Series Behavioral Modeling via Semantic RL-Tuned LLMs — A Case Study in Mental Health” has been accepted to ACM IMWUT and will be presented at UbiComp 2027!
Passive sensing over long periods lets us predict health outcomes continuously. But the models often break down when moved to a new dataset: traditional ML overfits to quirks of the cohort it was trained on, and LLMs have trouble reasoning over long, messy time-series. TimeSRL is a two-stage LLM framework that fixes this with a semantic bottleneck. The model first summarizes raw sensor signals as high-level natural language, then predicts behavioral outcomes from those summaries alone. We train the whole pipeline end-to-end with reinforcement learning with verifiable rewards, so the model learns useful summaries without any hand-labeled intermediate steps.
📄 Check out the paper: https://arxiv.org/abs/2605.21295