AI Refines Research Queries for Personalized Results.

Deep research agents rely on user requests as specifications to guide their evidence seeking and synthesis processes. For personalized deep research, these specifications must also reflect individual user goals, constraints, preferences, and evaluation criteria. This personalization can occur either within the research pipeline itself or by refining the initial user request before it enters the agent. This research focuses on the latter approach, introducing a system called G-STEER. G-STEER addresses the complex decisions involved in refining a query: identifying relevant framing factors, assessing if available user context supports them, and deciding whether to retrieve user memory, ask the user for more information, or stop and refine the query. To facilitate training, G-STEER organizes these framing factors as elicitation targets within an Intent Elicitation Graph, which maps out their dependencies. It then learns a clarification policy from graph-scaffolded trajectories that cover diverse factor dependencies and evidence conditions. This policy generates a refined query, balancing comprehensive target coverage with the costs associated with acquiring evidence. Experiments demonstrate that G-STEER achieves superior weighted target coverage and higher downstream report personalization across different deep research agents, all while significantly reducing the number of user questions compared to existing baselines.