DR³-Eval is a realistic, reproducible, and multimodal evaluation benchmark for Deep Research Agents, focusing on multi-file report generation tasks.
Existing benchmarks face a fundamental tension between realism, controllability, and reproducibility when evaluating deep research agents. DR³-Eval addresses this through the following design:… See the full description on the dataset page:
https://huggingface.co/datasets/NJU-LINK/DR3-Eval.
