# LifeAgentBench

> LifeAgentBench is a question-answering benchmark for evaluating whether large language models can reason over long-term, multi-dimensional lifestyle records. It covers long-horizon, cross-dimensional, and multi-user reasoning over diet, sleep, physical activity, and emotion data. LifeAgent is a separate tool-augmented baseline introduced for the benchmark.

## Key Facts

- Canonical title: LifeAgentBench: Benchmarking LLMs for Long-Horizon, Cross-Dimensional Lifestyle Health Reasoning
- Scale: 22,573 QA pairs
- User scope: 13,452 single-user questions and 9,121 multi-user questions
- Evaluated models: 13 representative LLMs in the 2026 camera-ready paper
- Source records: AI4FoodDB records from 100 anonymized participants
- Lifestyle dimensions: diet, sleep, physical activity, and emotion
- Reasoning types: Fact Query, Aggregated Statistics, Numeric Comparison, Conditional Query, and Trend Analysis
- Reasoning scopes: long-horizon, cross-dimensional, single-user, and multi-user
- Answer formats: Yes/No, scalar numerical values, short text, pairwise outputs, and multi-item lists
- Ground truth: deterministic answers produced through executable programs or database queries, with programmatic checks and manual inspection
- Baselines: Context Prompting and Database-augmented Prompting
- Primary metric: normalized answer accuracy; the paper defines exact-match and numeric-tolerance rules
- Intended use: research on evidence-grounded lifestyle reasoning, not diagnosis, treatment, medical decision making, or professional medical advice

## Key Findings

- Reliable evidence retrieval is a major bottleneck. Under Database-augmented Prompting, complete-evidence retrieval averages 28.79% across the 13 models, although nine models exceed 70% answer accuracy when complete evidence is retrieved.
- Performance degrades on aggregation-intensive questions, multi-item outputs, all-dimension questions, and multi-user reasoning.
- On the paper's most challenging subsets with the same Qwen2.5-7B backbone, LifeAgent reaches 40.16% average accuracy, compared with 7.74% for Context Prompting and 9.43% for Database-augmented Prompting.
- LifeAgent's decomposition, retrieval tools, and computation tools 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.

## Benchmark Results

- DeepSeek-V4-Pro is the strongest reported model, with 58.89% accuracy under Context Prompting and 55.67% under Database-augmented Prompting.
- Complete-evidence retrieval under Database-augmented Prompting averages 28.79% across the 13 models, but nine models exceed 70% final-answer accuracy when complete evidence is retrieved.
- GPT-4o illustrates the benchmark's complexity gap under Context Prompting: accuracy is 69.8% on Fact Query versus 15.6% on Aggregated Statistics, and 75.1% on Yes/No answers versus 28.1% on multi-item answers.

## LifeAgent

LifeAgent is a tool-augmented reasoning baseline, not the benchmark itself. It decomposes a user query into executable sub-questions, iteratively retrieves structured evidence, applies cohort-level aggregation and deterministic computation tools, and synthesizes an evidence-grounded response.

## Resources

- Project Page: https://gdfwj.github.io/LifeAgentBench/
- Camera-ready Paper PDF: https://gdfwj.github.io/LifeAgentBench/static/lifeagentbench-paper.pdf
- arXiv: https://arxiv.org/abs/2601.13880
- Dataset: https://huggingface.co/datasets/gdfwj/LifeAgentBench
- Code: https://github.com/gdfwj/LifeAgentBench
- Markdown Summary: https://gdfwj.github.io/LifeAgentBench/summary.md

## Citation

Tian, Ye; Wang, Zihao; Gungor, Onat; Fan, Xiaoran; and Rosing, Tajana. “LifeAgentBench: Benchmarking LLMs for Long-Horizon, Cross-Dimensional Lifestyle Health Reasoning.” EMNLP 2026.