Paperdoi:10.48550/arXiv.2608.27219
BALMS: benchmarking agentic LLMs for longitudinal mental health sensing
arXiv 2608.27219 (2026-08-27)
(opens in a new tab)Wu, Yu Yvonne et al., "BALMS: Benchmarking Agentic LLMs for Longitudinal Mental Health Sensing", arXiv:2608.27219, CC BY 4.0
TL;DR
First benchmark of LLM agentic systems for longitudinal mental-health sensing: 3 real-world longitudinal datasets, 2 task families (closed-form wellbeing-score prediction; rationale generation graded by LLM-as-judge), 3 agentic paradigms × 5 LLM backbones. Zero-shot agents rarely beat a simple mean baseline except with stronger backbones or compact, semantically meaningful features; chain-of-thought helps reasoning backbones but not temporal grounding or numerical correctness.
- Form factor: consumer wearable / phone
Abstract
Mental health assessment relies on episodic self-report scales, which convert subjective states such as stress into numerical scores but provide only sparse snapshots of wellbeing. Wearable devices offer longitudinal behavioral and physiological signals for continuous, low-burden monitoring. Recent LLM-driven personal-health agents enable natural language queries over wearable signals, but mainly handle short-term, retrieval-based lookups (e.g., highest step count over a week). They do not evaluate whether agents can reason over long-term signals to predict wellbeing scores paired with evidence-grounded rationales. To address this gap, we introduce BALMS, the first systematic benchmark of LLM-based agentic systems for longitudinal mental health sensing. BALMS spans 3 real-world longitudinal datasets, 2 task families (closed-form wellbeing-score prediction and rationale generation auto-graded by an LLM-as-Judge), 3 agentic paradigms evaluated across 5 open- and closed-source LLM backbones. We find that zero-shot agents rarely outperform a simple mean baseline, except with stronger backbones or compact, semantically meaningful features. Chain-of-thought prompting improves reasoning-oriented backbones, but does not guarantee temporal grounding or numerical correctness. Together with more analysis on efficiency and temporal scaling, BALMS highlights the need for longitudinal mental health agents that selectively retrieve history, ground temporal evidence, and reason over interpretable behavioral features.
Verbatim from DataCite (opens in a new tab) · Licence: https://creativecommons.org/licenses/by/4.0/legalcode
Results on ChatPPG leaderboards
All leaderboardsNo results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.
Tasks & signals
Signals
Devices
Keywords
Datasets used
How it works
Cite this paper
Wu, Y. Y., Pillai, A., Chen, Y., Zhang, Y., Regmi, S., Griffin, T. Z., Heinz, M. V., Marsch, L. A., Jacobson, N. C., & Campbell, A. (2026). BALMS: benchmarking agentic LLMs for longitudinal mental health sensing. arXiv 2608.27219. https://doi.org/10.48550/arXiv.2608.27219
Wu, Yu Yvonne, Arvind Pillai, Yuliang Chen, Yuwei Zhang, Sudarshan Regmi, Tess Z. Griffin, Michael V. Heinz, Lisa A. Marsch, Nicholas C. Jacobson, and Andrew Campbell. 2026. "BALMS: benchmarking agentic LLMs for longitudinal mental health sensing." arXiv 2608.27219. https://doi.org/10.48550/arXiv.2608.27219.
Wu, Y. Y., Pillai, A., Chen, Y., Zhang, Y., Regmi, S., Griffin, T. Z., Heinz, M. V., Marsch, L. A., Jacobson, N. C. and Campbell, A. (2026) 'BALMS: benchmarking agentic LLMs for longitudinal mental health sensing', arXiv 2608.27219. doi: 10.48550/arXiv.2608.27219.
Wu, Yu Yvonne, et al. "BALMS: benchmarking agentic LLMs for longitudinal mental health sensing." arXiv 2608.27219, 2026. https://doi.org/10.48550/arXiv.2608.27219.
Y. Y. Wu et al., "BALMS: benchmarking agentic LLMs for longitudinal mental health sensing," arXiv 2608.27219, 2026, doi: 10.48550/arXiv.2608.27219.
@misc{wu2026balms,
title = {{BALMS: benchmarking agentic LLMs for longitudinal mental health sensing}},
author = {Wu, Yu Yvonne and Pillai, Arvind and Chen, Yuliang and Zhang, Yuwei and Regmi, Sudarshan and Griffin, Tess Z. and Heinz, Michael V. and Marsch, Lisa A. and Jacobson, Nicholas C. and Campbell, Andrew},
year = {2026},
howpublished = {arXiv 2608.27219},
doi = {10.48550/arXiv.2608.27219}
}Added to the ChatPPG research hub
