The training recipe combines rubric-tree-based task synthesis, structured context management, and a three-stage pipeline spanning mid-training, supervised fine-tuning, and reinforcement learning. The project emphasizes releasing models, data, data synthesis scripts, and training code.
QUEST is useful for researchers and developers building deep research agents that need to search, cite, and synthesize long-form reports. Its benchmark comparisons focus on broad research-agent capabilities rather than a single narrow search task.
