What they did

The authors assembled BALMS from three real-world longitudinal datasets. It covers two task families: predicting numerical wellbeing scores and generating evidence-based rationales, with the latter automatically assessed by an LLM judge.

They compared three agentic paradigms across five open- and closed-source LLM backbones. Experiments examined zero-shot performance, chain-of-thought prompting, computational efficiency, the length of historical context, and different representations of wearable-derived features.

Key findings

• Zero-shot agents rarely outperformed a baseline that simply predicts the mean wellbeing score.
• Better results appeared with stronger LLM backbones or compact features carrying clearer semantic meaning.
• Chain-of-thought prompting improved some reasoning-oriented models, but did not consistently produce numerically correct predictions.
• Generated explanations could still lack grounding in the timing and sequence of the underlying wearable evidence.

Why it matters

The results challenge the assumption that general-purpose LLM agents can directly turn extensive wearable histories into dependable mental-health assessments. The benchmark points toward systems that retrieve only relevant periods, use interpretable behavioral features, and explicitly verify temporal and numerical claims.

Caveats

BALMS evaluates benchmark performance rather than clinical safety, diagnostic accuracy, or patient outcomes. Rationale quality is graded by another LLM, which may introduce evaluator bias, and the reported limitations suggest that these agents are not ready to replace validated assessments or professional judgment.