AIAny
Icon for item

WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data

Provides a diagnostic benchmark of 4,084 multiple-choice questions derived from longitudinal wearable time series, blood biomarkers, and demographics of 200 users to evaluate LLM reasoning on real-world wearable data. Uses 16 question types across complementary axes and a dual-grounding framework combining literature findings with population statistics.

Introduction

Wearable data are long-horizon, noisy, and multimodal, yet most LLM benchmarks focus on text or short-horizon numeric tasks. WearableQA fills that gap by turning real users' longitudinal wearable records into a diagnostic testbed that stresses numerical, temporal, and cross-signal reasoning in realistic, device-noise conditions. The benchmark preserves authentic distributions (up to 500 days per user) and pairs physiological interpretation with computational operations over time series.

Key Findings
  • Scale and realism: 4,084 ten-option multiple-choice questions built from wearable time series, blood biomarkers, and demographics of 200 users (up to 500 days per user). This setup preserves device noise and inter-individual variability so performance reflects real-world difficulty.
  • Dimensioned taxonomy: 16 question types organized along two axes—data vs health reasoning (computation over measurements vs physiological interpretation) and single- vs cross-signal reasoning (individual-signal analysis vs multimodal integration). This makes failure modes diagnostically informative.
  • Dual-grounding for reliability: questions are constructed using literature-grounded physiological relationships and statistically validated population patterns, reducing spurious or dataset-specific artifacts while retaining real distributions.
  • Model differentiation: evaluation across 14 LLMs yields accuracies from about 19.6% to 72.9% (10% chance baseline), with most models below 60%, indicating the task remains challenging and useful for benchmarking progress.
Who it's for and tradeoffs

Great fit if you are developing or evaluating LLMs, agentic toolchains, or tool-augmented systems that must reason about longitudinal physiological and behavioral signals, and you need a realistic, diagnostic benchmark to probe numeric, temporal, and multimodal reasoning. It’s also useful for researchers studying robustness to device noise and inter-individual variability.

Look elsewhere if you need labeled clinical outcomes, causal inference across interventions, or regulatory-grade clinical validation—the benchmark targets reasoning ability over observational wearable records rather than providing guidelines for deployment in clinical care. The multiple-choice format favors scalable automatic evaluation but may not capture every nuance of open-ended clinical explanation.

Information

  • Websitearxiv.org
  • OrganizationsMeta, KAIST, Korea University
  • AuthorsJi Soo Lee, Xilun Chen, Pierce Chuang, Ashish Shenoy, Jason Wei, Dohwan Ko, Hyunwoo J. Kim, Benoit Corda
  • Published date2026/09/04

More Items

Evaluates a weak teacher's RL-induced policy shift on the student's own rollouts and amplifies verifier-supported updates so stronger models can learn from weaker supervisors and surpass them. It rescales only verifier-supported policy-gradient components to preserve optimization fixed points while accelerating learning, reducing student updates versus standard RL or distillation.

Explores a practical mechanism for recursive self-improvement by post-training LLMs: uses a routing harness to record agent executions and convert traces into curriculum-guided supervised fine-tuning and on-policy distillation data, closing an evaluation-selection-update loop and improving benchmark performance.

Combines sparse verifier outcomes with dense privileged‑hindsight token scoring to learn an outcome‑calibrated, normalized distribution over complete responses for on‑policy self‑improvement. Key features: sign‑gated guidance (retain/reverse/disable per verifier advantage), profiled trajectory balance with one log‑partition per rollout group, and explicit correction against false‑positive self‑guidance.