BALMS: Benchmarking Agentic LLMs for Longitudinal Mental Health Sensing

Yu Yvonne Wu; Arvind Pillai; Yuliang Chen; Yuwei Zhang; Sudarshan Regmi; Tess Z. Griffin; Michael V. Heinz; Lisa A. Marsch

cs.CLAug 27, 2026

BALMS: Benchmarking Agentic LLMs for Longitudinal Mental Health Sensing

Authors: Yu Yvonne Wu, Arvind Pillai, Yuliang Chen, Yuwei Zhang, Sudarshan Regmi, Tess Z. Griffin, Michael V. Heinz, Lisa A. Marsch, +2 more

Organizations: Dartmouth College · University of Cambridge

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.

Figures & tables

Figure 1: LLM-based agents for longitudinal mental health sensing. Given a multi-channel wearable record and a user query, an agent combines retrieval, memory, tools, and reasoning to output a wellbeing score with an evidence-grounded rationale.

Figure 2: Token consumption on stress detection as wearable sensing duration grows from 7 to 28 days. Prompt-based input exceeds the model’s context limit.

Figure 3: Overview of BALMS , a benchmark for LLM-based agentic systems on longitudinal mental health sensing. Left: two task families share longitudinal passive sensing data as input. Right: the three agentic paradigms applied to both tasks. BALMS evaluates these paradigms across three real-world longitudinal datasets and five LLM backbones, scoring T1 by MAE and T2 by an LLM-as-a-Judge rubric.

Paradigm	System	Rationale	Raw passive sensing data	Tools	Memory	Dataset- centric
Prompt-based	Health-LLM ( Kim et al., 2024 )	✗	✗	✗	✗	✗
Prompt-based	Englhardt et al. ( Englhardt et al., 2024 )	✓	✗	✗	✗	✓
Memory-based	Chunk RAG ( Lewis et al., 2021 )	✗	✗	✗	✓	✗
Memory-based	RAPTOR ( Sarthi et al., 2024 )	✗	✗	✗	✓	✗
Tool-based	PHIA ( Merrill et al., 2026 )	✓	✗	✓	✗	✗
	GLOSS ( Choube et al., 2025 )	✓	✓	✓	✗	✓

Table 2: Comparison of representative LLM-based agentic systems for passive mobile sensing. Rationale = produces a free-form rationale alongside or instead of a closed-form label.

Backbone	Method	DiversityOne(Mood)		PMData(Stress)		GLOBEM(Anxiety)
Backbone	Method	MAE ↓	+CoT MAE ↓	MAE ↓	+CoT MAE ↓	MAE ↓	+CoT MAE ↓
Reference baseline	Mean Predictor	0.58		0.48		0.80
Open-source LLMs
Qwen2.5-7B-Instruct	Health-LLM	0.60	0.62	0.65	0.55	1.17	1.30
	RAG	0.96	0.99	0.72	0.63	1.20	1.37
	Raptor	1.15	1.29	0.78	0.64	1.21	1.22

Table 3: Zero-shot mental health detection. We compare 4 agentic systems with 5 backbones reporting MAE without and with CoT. No paradigm wins across all three datasets. Best per column in bold , second-best underlined .

Figure 4: LLM-as-Judge results on GLOBEM. Each subplot corresponds to one backbone, overlaying 3 agentic systems; GQ is the Tier-2 general-quality overall Likert score, rescaled to

[0,1]

for comparability. Higher is better on all axes. Only Claude-Haiku-4.5 closes both the rationale-quality and temporal-grounding gaps, while open-source instruct backbones produce fluent rationales but struggle with temporal reasoning and rarely invoke alignment (C1) or structure (C5).

Figure 5: Effect of look-back window length on MAE for Qwen2.5-14B across PMData and GLOBEM. Corresponding latency curves are reported in Figure 7 . Health-LLM degrades as the window grows, RAG benefits the most, and PHIA stays stable but remains schema-sensitive on GLOBEM.

Figure 6: Average per-sample latency of each agentic system with Qwen2.5-14B-Instruct.

Backbone	Method	Full		Fitbit
Backbone	Method	MAE ↓	+CoT ↓	MAE ↓	+CoT ↓
Qwen2.5- 7B-Instruct	Health-LLM	1.17	1.30	1.12	1.24
	RAG	1.20	1.37	1.43	1.27
	Raptor	1.21	1.22	1.22	1.20
	PHIA	1.48	1.54	1.18	1.08
Mistral-7B- Instruct-v0.3	Health-LLM	1.83	2.00	1.31	1.85

Table 4: Sensor-set sensitivity on GLOBEM. Each agentic system on GLOBEM under two sensor configurations: Full (mobile + smartwatch) vs. Fitbit (smartwatch only), w/o and w/cot. Dropping raw mobile streams and keeping only Fitbit semantic rich features often matches or beats the full setup.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Task (Likert)	Look-back	# Samples
DiversityOne	Mood (1–5)	7 days	1,355
PMData	Stress (1–5)	14 days	1,448
GLOBEM	PHQ-4 Anxiety (0–3)	14 days	1,640

Appendix

Table 5: Processed evaluation set per dataset. For each dataset we list the prediction target, the default look-back window used for T1 main-table runs, and the number of (target-day, user) samples after cohort selection and missing-label filtering.

Backbone	System	Faithfulness	Evidence	Domain	Clarity	Overall
Qwen2.5-7B	Health-LLM	3.98	4.55	4.29	4.79	4.11
	RAG	3.92	3.24	4.25	4.55	3.96
	PHIA	4.07	3.50	4.43	4.71	4.13
Mistral-7B-Instruct-v0.3	Health-LLM	3.08	4.61	4.45	4.65	4.06
	RAG	1.46	4.03	4.43	4.70	3.81
	PHIA	2.98	3.05	4.05	4.30	3.59

Appendix

Table 6: Tier-2 (general-quality) Likert scores on GLOBEM, 1–5 scale. Each row is one (backbone, agentic system) cell. Safety is omitted as every cell scores

5.0

. Best per column in bold ; second-best underlined .

Dataset	Modal label	Mode rate	MAE
GLOBEM (anxiety)	2	85.9% (1409/1640)	1.52
PMData (stress)	3	77.4% (1132/1463)	0.57
DiversityOne (mood)	None	99.5% (757/761)	2.11

Appendix

Table 7: PHIA mode-collapse rates with Qwen2.5-14B-Instruct (seed 0). The modal label is the single most-frequent prediction across all held-out windows. “None” on DiversityOne indicates a non-parseable final answer.

Panel of Figure 7 — Figure 7: Average per-sample latency as a function of look-back window length for Qwen2.5-14B on PMData and GLOBEM.

Figure 8: Average per-sample latency of each agentic system with Qwen2.5-7B-Instruct.

Figure 9: Average per-sample latency of each agentic system with Mistral-7B-Instruct-v0.3.

Figure 10: Average per-sample latency of each agentic system with DeepSeek-R1-Distill-Qwen-14B.

Claim	Qwen2.5-7B	Claude-Haiku-4.5
Sensor + value match	4/4	4/4
Correct day attribution	2/4	4/4
Multi-day trend cited	0	1 (loc. entropy $1.80{\to}0.85$ )
Wrong-day errors	“139 min sleep on day 2” (actually day 3); “334 s screen on day 10” (actually day 12)	—
Cherry-picked single-day deltas	“sleep efficiency drops day 7 $\to$ 8” (rest of window contradicts)	—

Appendix

Table 8: Per-claim verification of numeric assertions in the two rationales. Qwen’s two wrong-day attributions cite values that exist elsewhere in the window but on the wrong day — the failure mode the evidence_grounding axis targets.

Explore similar work

LifeAgentBench: Benchmarking LLMs for Long-Horizon, Cross-Dimensional Lifestyle Health Reasoning
Jan 20, 2026Ye Tian, Zihao Wang, Onat Gungor +2Healthcare Long-Horizon Agent Evaluation
University of California San Diego, La Jolla, CA, USA · Google Research, Mountain View, CA, USA
AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications
Feb 26, 2026Yujie Zhao, Boqin Yuan, Junbo Huang +9Agent Memory Long-Horizon Agent Evaluation
UCSD · Recursive.
TimeSRL: Generalizable Time-Series Behavioral Modeling via Semantic RL-Tuned LLMs -- A Case Study in Mental Health
May 20, 2026Yuang Fan, Lilin Xu, Millie Wu +8Mental Health Domain Generalization
Columbia University, USA · University of North Carolina at Chapel Hill, USA · Yale University, USA +3

Jan 20, 2026J/K move · Enter open · S save

LifeAgentBench: Benchmarking LLMs for Long-Horizon, Cross-Dimensional Lifestyle Health Reasoning

Ye Tian, Zihao Wang, Onat Gungor, Xiaoran Fan+1

University of California San Diego, La Jolla, CA, USA · Google Research, Mountain View, CA, USA

Healthcare Long-Horizon Agent Evaluation AI Agent Benchmarks LLM Tool Use

Open paper AskPDF arXiv

Personalized lifestyle health analysis requires long-horizon, multi-dimensional reasoning over heterogeneous lifestyle signals, and recent advances in mobile sensing and large language models (LLMs) make such support increasingly feasible. However, the capabilities of current LLMs in this setting remain insufficiently understood due to the lack of systematic benchmarks. In this paper, we introduce LifeAgentBench, a large-scale QA benchmark for long-horizon, cross-dimensional, and multi-user lifestyle health reasoning, containing 22,573 questions spanning from basic retrieval to complex reasoning. We release an extensible benchmark construction pipeline and a standardized evaluation protocol, deriving verifiable answers through executable queries and programs to support reliable assessment. We then systematically evaluate 13 representative LLMs on LifeAgentBench and identify key bottlenecks in long-horizon aggregation and cross-dimensional reasoning. Motivated by these findings, we propose LifeAgent, a tool-augmented reasoning baseline that decomposes complex queries, performs multi-step evidence retrieval, and invokes tools for deterministic aggregation. LifeAgent substantially enhances LLMs' capabilities on challenging reasoning tasks, achieving clear improvements over widely used baselines and showing potential for health reasoning in everyday scenarios. The benchmark is publicly available.

Abstract

Explore similar work

LifeAgentBench: Benchmarking LLMs for Long-Horizon, Cross-Dimensional Lifestyle Health Reasoning

AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications

TimeSRL: Generalizable Time-Series Behavioral Modeling via Semantic RL-Tuned LLMs -- A Case Study in Mental Health