QUEST: Training Deep-Research Agents with Synthetic Tasks

Translation status: The English translation is pending. The section structure is synchronized with the Chinese source.

Motivation: Unifying Fact Seeking, Citation Grounding, and Report Synthesis

Method: Rubric-Tree Synthesis, Context Compression, and Three-Stage Training

Data Synthesis with a Unified Rubric Tree

Context Management

Training Pipeline

Infrastructure

Results

Evaluation Setup

Main Results

Training-Stage and Model-Scale Ablations

Unsuccessful Attempts

Position in the Deep-Research Landscape