QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks

In three linesQUEST is a family of open-source models (2B to 35B) trained as deep research agents via data synthesis pipeline and RL. With only 8K synthetic tasks, QUEST matches or exceeds proprietary systems across 8 research benchmarks, excels at citation grounding and report synthesis. Models, data, and training scripts released.

Summary generated by Claude — human-verified