InterDeepResearch: Enabling Human-Agent Collaborative Information Seeking through Interactive Deep Research
New system from Bo Pan's team tackles the 'black box' problem in AI research agents, letting users steer the process.
A research team led by Bo Pan has published a paper on arXiv introducing InterDeepResearch, a system designed to solve a critical flaw in current AI-powered research tools. Most existing systems, like those using GPT-4 or Claude 3.5 agents, operate as autonomous 'query-to-report' black boxes, leaving users as passive recipients. This approach fails to incorporate a user's personal insights, contextual knowledge, or evolving research goals, limiting the depth and relevance of the final output.
InterDeepResearch addresses this by creating a framework for true human-agent collaboration. Its core is a dedicated research context management system that organizes information into a three-level hierarchical architecture (information, actions, and sessions). This structure enables dynamic context reduction to prevent LLM context window exhaustion and allows for cross-action backtracing to verify evidence provenance. The user interface provides three coordinated visual views for sensemaking and dedicated mechanisms for interactively navigating the research context.
The system was evaluated on the Xbench-DeepSearch-v1 and Seal-0 benchmarks, where it achieved performance competitive with state-of-the-art autonomous deep research systems. More importantly, a formal user study demonstrated its superior effectiveness in supporting collaborative information seeking. This represents a significant shift from fully automated agents to interactive tools where human expertise guides and refines the AI's search and synthesis process.
- Solves the 'black box' problem by enabling real-time user steering and process observability in AI research.
- Uses a novel hierarchical context framework to manage information, actions, and sessions, preventing LLM context exhaustion.
- Achieved competitive benchmark performance while a user study proved its effectiveness for human-AI collaboration.
Why It Matters
It transforms AI from an autonomous researcher into a collaborative partner, making deep research more transparent, steerable, and integrated with human expertise.