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

Ye Tian; Zihao Wang; Onat Gungor; Xiaoran Fan; Tajana Rosing

At a glance

LifeAgentBench at a Glance

LifeAgentBench is a question-answering benchmark for evaluating whether LLMs can reason over long-term, multi-dimensional lifestyle records, including long-horizon, cross-dimensional, and multi-user reasoning.

Benchmark scale: 22,573QA pairs
Model coverage: 13evaluated LLMs
Record coverage: 4lifestyle dimensions

The four dimensions are diet, sleep, physical activity, and emotion. The benchmark contains 13,452 single-user and 9,121 multi-user questions.

Research gap

Why LifeAgentBench?

01
Existing evaluation
Health and lifestyle QA benchmarks often focus on one domain, short temporal contexts, or task-specific applications.
02
Real lifestyle records
Longitudinal records require aggregation over time, alignment across heterogeneous signals, and comparison across users.
03
LifeAgentBench
The benchmark tests verifiable reasoning across long time windows, multiple lifestyle dimensions, and single- or multi-user settings.

LifeAgentBench studies a gap that conventional health and lifestyle QA benchmarks often leave underexplored: reasoning over longitudinal, heterogeneous lifestyle records across multiple health-related dimensions and users.

Verifiable benchmark construction

Benchmark Composition and Construction

LifeAgentBench is built from anonymized AI4FoodDB records for 100 participants. Records are normalized into relational tables, questions are instantiated from reusable templates, and deterministic answers are produced by executable programs and queries.

Normalize recordsAlign diet, sleep, activity, and emotion by participant ID and timestamp.
Instantiate questionsCompose time-window, aggregation, comparison, condition, and trend operators.
Execute ground truthRun deterministic programs or database queries for verifiable answers.
Check qualityApply programmatic checks and manual inspection.

LifeAgentBench composition showing 13,452 single-user and 9,121 multi-user questions, vocabulary, five task categories, and answer formats — Benchmark composition: user scope and domains (left), question vocabulary (center), and task and answer-format distribution (right).

Benchmark taxonomy

What Does LifeAgentBench Test?

The official taxonomy contains five reasoning types. Questions additionally vary by time horizon, domain scope, user scope, and answer format.

FQ Fact QueryAS Aggregated StatisticsNC Numeric ComparisonCQ Conditional QueryTA Trend Analysis

The questions below are lightly normalized from released benchmark examples. Placeholders replace participant identifiers, dates, and numeric values; answers are intentionally omitted.

Fact Query · FQ

“How many moderately active minutes did {participant_id} record on {date}?”

Retrieves one atomic fact for a specified participant and date.

Aggregated Statistics · AS

“What was {participant_id}’s total cardio activity time within one week, starting {start_date}?”

Aggregates a measure over a longitudinal window.

Conditional Query · CQ

“Over one week, on how many days did {participant_id} exceed both {step_threshold} steps and {sleep_threshold} minutes asleep?”

Applies thresholds while aligning activity and sleep records by date.

Numeric Comparison · NC

“Which duration was longest for {participant_id} on {date}: light, moderate, or very active minutes?”

Compares multiple numerical measurements and returns the matching label.

Trend Analysis · TA

“Did average steps and average deep-sleep minutes both increase consecutively across all participants over {number_of_days} days?”

Combines multi-user aggregation, cross-dimensional alignment, and consecutive trend reasoning.

Experimental takeaways

Key Findings

Finding 1

Reliable evidence retrieval remains a key bottleneck.

Across the 13 models, complete-evidence retrieval under Database-augmented Prompting averages only 28.79%. When complete evidence is retrieved, however, nine models exceed 70% final-answer accuracy.

Finding 2

Aggregation, structured outputs, and larger evidence scopes remain difficult.

Models struggle most with aggregation-intensive reasoning and structured outputs, and performance further degrades as questions span more lifestyle dimensions and users.

Finding 3

LifeAgent strengthens reasoning on the most challenging subsets.

With the same Qwen2.5-7B backbone, LifeAgent reaches 40.16% average accuracy on the hardest subsets, compared with 7.74% for Context Prompting and 9.43% for Database-augmented Prompting.

Finding 4

LifeAgent’s three core components make complementary contributions.

Removing decomposition reduces all-dimension accuracy from 52.75% to 38.76%, while removing retrieval tools reduces health-indicator accuracy from 69.84% to 33.84%. Removing computation tools also lowers performance.

Reported evaluation

Benchmark Results

Accuracy of 13 evaluated LLMs under Context Prompting and Database-augmented Prompting.
Model	Context Prompting Accuracy (%)	Database-augmented Prompting Accuracy (%)
Open-source LLMs
DeepSeek-Coder-1.3B	1.09	1.26
Llama-3.2-3B	20.18	13.47
Phi-3.5-mini-3.8B	20.57	16.16
Mistral-v0.3-7B	30.97	9.03
Qwen-2.5-7B	40.45	21.45
Llama-3.1-8B	20.65	21.53
Gemma-2-9B	24.44	14.54
Qwen3.5-9B	45.55	25.76
Llama-3.1-70B	40.51	13.91
Closed-source LLMs
Gemini 2.5 Flash-Lite	44.81	39.04
Claude-3-Haiku	35.30	29.30
GPT-4o	57.02	34.71
DeepSeek-V4-Pro	58.89	55.67

Context Prompting places pre-filtered records in the model context. Database-augmented Prompting asks the model to generate a read-only SQL query, executes it, and returns the evidence for final answering. See the paper for exact answer normalization and numeric-tolerance rules.

Scatter plot comparing Context Prompting and Database-augmented Prompting accuracy for 13 evaluated language models — Overall accuracy of the 13 evaluated LLMs under Context Prompting and Database-augmented Prompting.

GPT-4o accuracy by LifeAgentBench question type and answer format under Context Prompting, Database-augmented Prompting, and Database-augmented Prompting conditioned on complete evidence retrieval — GPT-4o performance by question type and answer format, including final-answer accuracy conditioned on complete evidence retrieval.

Baseline system

LifeAgent: A Tool-Augmented Reasoning Baseline

LifeAgent is not the benchmark itself. It is a separate baseline introduced to address the evidence-retrieval, aggregation, and cross-dimensional reasoning failures exposed by LifeAgentBench.

LifeAgent framework showing user-query decomposition, an iterative thought-action-observation loop with retrieval and computation tools, evidence-grounded synthesis, and response generation — LifeAgent decomposes a user query, iteratively retrieves and computes evidence with tools, and synthesizes an evidence-grounded response.

1User queryReceive a lifestyle health question.
2In-context decompositionIdentify parameters, sub-questions, and an executable plan.
3Iterative tool useRun a thought–action–observation loop and update evidence.
4Evidence-grounded synthesisIntegrate records and computed intermediate results.
5ResponseProduce an answer grounded in accumulated evidence.

Structured data retrieval

Event lookups and daily aggregated time series constrained by domain, time window, and granularity.

Cohort-level aggregation

Multi-user summaries, ranking by selected metrics, and criterion-based group comparisons.

Deterministic computation

Arithmetic, aggregation, threshold checks, comparisons, consecutiveness, and trend analysis.

Why LifeAgent matters: the results show that explicit decomposition, iterative evidence retrieval, and deterministic computation can strengthen LLM reasoning on the benchmark’s most challenging tasks.

Quick start

Use LifeAgentBench

The repository supports Context Prompting and Database-augmented Prompting. This concise path uses the documented Context Prompting workflow.

01
Download the benchmark
Clone the repository with Git LFS enabled, or download the dataset from Hugging Face.
```
git lfs install
```

02

Run a model

Run the documented Context Prompting evaluator on the released JSONL files.

python eval_simple.py \
  --data-root ./gen_data_processed/simple \
  --eval-root ./eval \
  --model meta-llama/Llama-3.2-3B-Instruct \
  --max-new-tokens 32 \
  --api-key ""

03
Evaluate predictions
Compute overall summary statistics from the generated outputs.
```
python stat_simple.py \
  eval/model_name/all_outputs.jsonl
```

See the full README for environment setup, output format, Database-augmented Prompting, and dataset-construction instructions.

Paper

Abstract

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.

Common questions

FAQ

What is LifeAgentBench?

LifeAgentBench is a 22,573-question benchmark for evaluating long-horizon, cross-dimensional, and multi-user reasoning over longitudinal lifestyle records.

What types of reasoning does LifeAgentBench evaluate?

It evaluates five reasoning types—Fact Query, Aggregated Statistics, Numeric Comparison, Conditional Query, and Trend Analysis—across different time horizons, lifestyle dimensions, user scopes, and answer formats.

What lifestyle dimensions are included?

LifeAgentBench includes diet, sleep, physical activity, and emotion records derived from AI4FoodDB.

How are ground-truth answers generated?

Each question is mapped to an executable program or database query that produces a deterministic answer. The pipeline also applies programmatic checks and manual inspection.

How is LifeAgentBench different from conventional health QA benchmarks?

LifeAgentBench emphasizes reasoning over longitudinal, heterogeneous records, including cross-dimensional integration and multi-user comparison, rather than limiting evaluation to isolated questions, one domain, or short contexts.

What is LifeAgent?

LifeAgent is a tool-augmented reasoning baseline for the benchmark. It decomposes complex queries, retrieves evidence iteratively, uses deterministic computation tools, and synthesizes an evidence-grounded response.

How can I evaluate a new LLM on LifeAgentBench?

Download the released JSONL data, run either the Context Prompting or Database-augmented Prompting evaluation script, and compute summary accuracy with the corresponding statistics script.

Is LifeAgentBench intended to evaluate clinical diagnosis or clinical decision making?

No. LifeAgentBench and LifeAgent are research resources for evaluating evidence-grounded lifestyle reasoning; they are not intended for diagnosis, treatment, medical decision making, or professional medical advice.

Reference

Citation

@article{tian2026lifeagentbench,
  title={LifeAgentBench: A Multi-dimensional Benchmark and Agent for Personal Health Assistants in Digital Health},
  author={Tian, Ye and Wang, Zihao and Gungor, Onat and Fan, Xiaoran and Rosing, Tajana},
  journal={arXiv preprint arXiv:2601.13880},
  year={2026}
}

Paper PDF arXiv Markdown summary llms.txt