The authors introduce BALMS, the first systematic for evaluating LLM-based agentic systems on longitudinal mental health sensing using wearable data. The benchmark spans three real-world datasets, two task families (wellbeing-score prediction and rationale generation), three agentic paradigms, and five open- and closed-source LLM backbones.
Key findings show that zero-shot agents rarely outperform a simple mean baseline, except when using stronger backbones or compact, semantically meaningful features. Chain-of-thought prompting improves reasoning-oriented backbones but does not guarantee temporal grounding or numerical correctness. The results highlight the need for agents that selectively retrieve history, ground temporal evidence, and reason over interpretable behavioral features, making this benchmark valuable for researchers and developers working on wearable-based mental health assessment.