POISE-Bench

Do LLM Agents Know What to Look Up? Certified Information Seeking under Partial Observability and a Budget

Shahriar Shakir Sumit, Murat Tahtali, Seyedali Mirjalili and Md Tauhidul Islam. Submitted to the Journal of Data-centric Machine Learning Research (DMLR), 2026.

This is the data of POISE-Bench (Partially Observable Information-Seeking Evaluation). The code that builds, runs and scores it is on GitHub at github.com/ShahriarShakirSumit/POISE-Bench.

What is here

Each question comes from a small hidden world, so we know exactly which records settle the answer and what the cheapest way to find them is. An agent opens hidden records one at a time within a budget and then answers or says that the records cannot decide.

The folder data/v2 holds the 8,000 training, 1,600 development and 2,000 test questions, one JSON list per split, together with the report of the checks every question passed. data/naturalistic has 78 human-curated tasks from areas such as medicine, law and engineering. data/musique lists the 600 MuSiQue questions of our real-text study by their ids, since the questions themselves come from the official MuSiQue release. The folder results holds every answer the 14 models gave, with the full turn-by-turn record of each episode, and annotation holds the items shown to our two human annotators, the answer key and their sheets.

In each question, the agent may see question, answer_space, instructions, initial_observation and, for every hidden record, its id and title; the text of a record appears only after the agent opens it. The fields gold_answer, minimal_sufficient_evidence, oracle_cost and budget_tiers are for scoring, and _world keeps the hidden world for analysis, so it should never be shown to an agent.

How to use it

Clone the GitHub repository and fetch this data into the same folder, and every script there will find it.

git clone https://github.com/ShahriarShakirSumit/POISE-Bench.git
cd POISE-Bench
pip install -r requirements.txt
python -c "from huggingface_hub import snapshot_download; snapshot_download('Shahriar10/POISE-Bench', repo_type='dataset', local_dir='.', allow_patterns=['data/*', 'results/*', 'annotation/*'])"

Licence and citation

The data is released under the MIT licence. The MuSiQue questions and paragraphs that appear in the model records belong to their authors and keep their CC BY 4.0 licence. Please do not train on the test split, so that results stay comparable. If you use POISE-Bench, please cite our paper.

@misc{sumit2026poisebench,
  title  = {Do {LLM} Agents Know What to Look Up? Certified Information Seeking under Partial Observability and a Budget},
  author = {Sumit, Shahriar Shakir and Tahtali, Murat and Mirjalili, Seyedali and Islam, Md Tauhidul},
  year   = {2026},
  note   = {Submitted to the Journal of Data-centric Machine Learning Research}
}