| --- |
| license: other |
| license_name: deer-dataset-license |
| license_link: LICENSE |
| language: |
| - en |
| tags: |
| - deep-research |
| - benchmark |
| - rubric-based-evaluation |
| - agent |
| - fact-verification |
| pretty_name: DEER |
| size_categories: |
| - n<1K |
| --- |
| # DEER: A Benchmark for Evaluating Deep Research Agents on Expert Report Generation |
|
|
| 📄 **Paper:** [OpenReview](https://openreview.net/forum?id=ILRx5neJk6) |
| 🏆 **Accepted at ICML 2026** |
|
|
| Existing evaluations for Deep Research Agents often remain surface-level when assessing long-form expert reports. DEER enables evaluation from an expert perspective using 101 fine-grained rubrics and task-specific human expert evaluation guidance. It also links every factual statement in a report to relevant external evidence and verifies whether the content is actually grounded, enabling a more complete assessment of report reliability. |
|
|
| ## Dataset Access |
|
|
| The complete dataset is distributed as a password-protected 7z archive. |
|
|
| **Archive:** `deer_dataset.7z` |
| **Password:** `0000` |
|
|
| ```bash |
| 7z x deer_dataset.7z |
| ``` |
|
|
| The archive is password-protected to prevent web-enabled agents from finding and using expert-written evaluation guidance that specifies the key content expected in reports for each DEER query. |
|
|
| ## Usage |
|
|
| DEER may be used for non-commercial research, evaluation, model training, fine-tuning, continued pretraining, and development. |
|
|
| To preserve the integrity of the benchmark: |
|
|
| * do not redistribute, mirror, republish, or publicly host the original dataset; |
| * do not upload the extracted dataset to public repositories, websites, dataset platforms, or file-sharing services; and |
| * direct other users to this official repository instead of sharing copies of the dataset. |
|
|
| Commercial use requires prior written permission from LG AI Research. |
|
|
| See the full [DEER Dataset License](https://huggingface.co/datasets/LG-AI-Research/DEER-Deep-Research-Benchmark/blob/main/LICENSE). |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{ |
| han2026deer, |
| title={{DEER}: A Benchmark for Evaluating Deep Research Agents on Expert Report Generation}, |
| author={Janghoon Han and Heegyu Kim and Changho Lee and Dahm Lee and Min Hyung Park and Hosung Song and Stanley Jungkyu Choi and Moontae Lee and Honglak Lee}, |
| booktitle={Forty-third International Conference on Machine Learning}, |
| year={2026}, |
| url={https://openreview.net/forum?id=ILRx5neJk6} |
| } |
| ``` |
|
|
| ## License |
|
|
| Copyright © 2026 LG AI Research. All rights reserved. |
|
|
| DEER is released under the custom [DEER Dataset License](https://huggingface.co/datasets/LG-AI-Research/DEER-Deep-Research-Benchmark/blob/main/LICENSE). |
|
|
