TimeSRL: Semantic RL-Tuned LLMs for Mental Health Time-Series
A new two-stage LLM framework, named TimeSRL, has been introduced by researchers for modeling time-series behavior, specifically in mental health forecasting. Initially, the model converts raw sensor data into high-level natural language, subsequently predicting outcomes based solely on these abstractions, thus establishing a semantic bottleneck that enhances generalization. It employs Group Relative Policy Optimization (GRPO) combined with Reinforcement Learning from Verifiable Rewards (RLVR) for end-to-end optimization, eliminating the need for gold-standard intermediate annotations. TimeSRL demonstrates superior performance on cross-dataset distribution shifts in longitudinal passive sensing data, surpassing both conventional ML and standard LLMs. This research is documented in arXiv preprint 2605.21295.
Key facts
- TimeSRL is a two-stage LLM framework for time-series behavioral modeling.
- It abstracts raw signals into high-level natural language before prediction.
- Optimized using GRPO with RLVR, no gold intermediate annotations needed.
- Achieves state-of-the-art performance on mental health prediction tasks.
- Addresses cross-dataset distribution shifts in longitudinal passive sensing.
- Published as arXiv:2605.21295.
- Traditional ML overfits cohort-specific artifacts; LLMs struggle with heterogeneous time-series.
- Semantic bottleneck forces reasoning over generalizable concepts.
Entities
Institutions
- arXiv