AREAs-Lab

AREAs-Lab is a synthetic benchmark for studying how AI assistants turn incomplete user requests into actionable requirements for AI systems. It evaluates an agent's ability to uncover missing goals, preferences, constraints, and edge cases by combining targeted clarification questions with inspection of the underlying data.

The benchmark reported in our paper contains 151 curated tasks grounded in 16 public datasets, spanning domains such as medicine, law, finance, education, scientific research, and media. Each instance connects a source dataset with a synthetic user persona, a complete reference requirement, and an intentionally underspecified initial request. The requirements reflect both what the user wants and what the data demands, including structural patterns, formatting artifacts, and domain-specific constraints.

Used with the accompanying AREAs-Lab environment, these instances support interactive evaluation with an AI-simulated user that reveals relevant information when appropriately prompted. The assistant iteratively refines its specification, which is evaluated against the reference using atomic requirement precision, recall, and F1. AREAs-Lab supports research on clarification strategies, user intent understanding, data-grounded reasoning, and human-AI collaboration.

Where to start

Start with data_synthesized/, where artifacts are organized by their upstream dataset identifiers. For example, the BillSum directory contains the following files:

File	Contents
synthesized_output.json	Synthesized personas, task requirements and summaries, validation results, and generation metadata.
data_analysis.json	Analysis of the source dataset's metadata and schema.
sample_analysis.json	Features extracted from sampled instances to ground task construction.
ground_truth_decompose.json	Decomposed reference requirements for atomic evaluation.

The repository also includes data_sampled/, organized by source dataset, with files such as FiscalNote/billsum/data_sampled.json.

The 151-task figure refers to the filtered benchmark described in the paper, rather than the number of files or all task-generation records. Use the paper's filtering protocol and accompanying code when reproducing the benchmark.

What's in a task

In the BillSum synthesis file, persona entries are stored under top-level keys such as user_1. Each entry contains user_info, tasks_info, and validation_results.

Component	Description
user_info	The persona's role, competencies, limitations, business motivation, and workflow challenges.
tasks_info[].task_name / difficulty	The task title and its intended difficulty level, Medium or High.
tasks_info[].task_requirement	The complete reference specification, including objectives, constraints, and task logic.
tasks_info[].elevator_pitch_summary	A short, intentionally incomplete request used to initialize elicitation.
tasks_info[].deep_dive_summary	A longer explanation of the task and its relationship to the data.
tasks_info[].data_columns, data_features, and dataset_alignment_explanation	The input fields and data characteristics that motivate the task.
validation_results	Model-generated quality assessments of the synthesized tasks.

These are nested JSON artifacts with synthesis metadata, rather than a flat table containing one row per benchmark task.

Load an example

Download a synthesis file with huggingface_hub and read it as JSON:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="ZihaoZhang/AREAs-Lab",
    repo_type="dataset",
    filename="data_synthesized/FiscalNote/billsum/synthesized_output.json",
)

with open(path, encoding="utf-8") as f:
    artifact = json.load(f)

for key, user in artifact.items():
    if not key.startswith("user_") or not isinstance(user, dict):
        continue
    for task in user.get("tasks_info", []):
        print(task["task_name"])
        print(task["elevator_pitch_summary"])

This example inspects the released synthesis records; it does not apply benchmark filtering or run an evaluation. See the code repository for the accompanying implementation.

Citation

If you use AREAs-Lab, please cite our paper and the relevant upstream datasets:

@misc{cai2026areaslab,
  title={AREAs-Lab: An Interactive Environment for AI-driven Requirement Elicitation for AI Systems},
  author={Pengshan Cai and Zihao Zhang and Ting Jin and Chenyang Zhu and Kushal Chawla and Sangwoo Cho and Scott Novotney and Yebowen Hu and Fei Liu and Shi-Xiong Zhang and Sambit Sahu},
  year={2026},
  eprint={2608.28979},
  archivePrefix={arXiv},
  primaryClass={cs.HC},
  url={https://arxiv.org/abs/2608.28979}
}