InterDeepResearch: Enabling Human-Agent Collaborative Information Seeking through Interactive Deep Research

Deep research systems have recently gained widespread application in academia and industry. These systems, equipped with LLM-based agents, can automatically perform information search, collection, and synthesis to address complex research goals posed by users. Traditional information retrieval systems typically rely on hard-coded retrieval workflows, while Retrieval-Augmented Generation (RAG) methods leverage LLMs' general intelligence to process intermediate information but still maintain rigid workflows. In contrast, deep research systems fully exploit the autonomy of LLM-based agents, enabling them to proactively plan, invoke information retrieval tools, and process information like human researchers, thereby accomplishing more complex and long-horizon information seeking tasks.

Despite significant advances, existing deep research systems are predominantly designed for a single-turn query-to-report paradigm, limiting users to a passive role without the opportunity to participate in the research process. However, substantial research in information retrieval has established that many information seeking tasks benefit significantly from human-in-the-loop collaboration, particularly when users need to contribute personal insights, contextual knowledge, and evolving research intents that cannot be fully captured in initial queries. This need for user involvement manifests in two ways during deep research: at a high level, users may need to adjust research strategies based on emerging insights; at a low level, users may need to intervene on critical execution details, such as correcting the agent when it relies on incorrect assumptions or specifying precise search keywords that require contextual knowledge.

InterDeepResearch introduces a dedicated research context management framework to support human-agent collaborative information seeking. This framework organizes research context into a hierarchical architecture with three levels: information, actions, and sessions. It enables dynamic context reduction to prevent LLM context exhaustion and cross-action backtracing for evidence provenance. The system interface integrates three coordinated views for visual sensemaking and provides dedicated interaction mechanisms for interactive research context navigation. This innovative system design is unprecedented in existing deep research systems.

• �� Research Context Management Framework: Organizes research context into a hierarchical architecture with information, actions, and sessions.
• �� Dynamic Context Reduction: Prevents LLM context exhaustion.
• �� Cross-action Backtracing: Traces conclusions back to supporting evidence.
• �� System Interface: Integrates three coordinated views for visual sensemaking.
• �� Interaction Mechanisms: Supports interactive research context navigation.

Evaluations on the Xbench-DeepSearch-v1 and Seal-0 benchmarks demonstrate that InterDeepResearch achieves competitive performance compared to state-of-the-art deep research systems. A formal user study shows that the system effectively supports human-agent collaborative information seeking, allowing users to better participate in and guide the research process. The system supports users to pause research, inject new requirements, adjust strategies, and redirect focus based on emerging insights without restarting the entire research cycle.

Evaluations on the Xbench-DeepSearch-v1 and Seal-0 benchmarks demonstrate that InterDeepResearch achieves competitive performance compared to state-of-the-art deep research systems. A formal user study shows that the system effectively supports human-agent collaborative information seeking, allowing users to better participate in and guide the research process. Users can flexibly intervene and guide the agent's research direction during the process. The system supports users to pause research, inject new requirements, adjust strategies, and redirect focus based on emerging insights without restarting the entire research cycle.

InterDeepResearch can be directly applied in scenarios such as academic research, market analysis, and domain-specific investigations. The system's dynamic context management and interaction mechanisms make it adaptable to different user needs, enhancing the efficiency and accuracy of information retrieval.

Despite significant advances, the system may face challenges in handling particularly complex or long-duration research tasks, as context management, despite dynamic reduction mechanisms, may still lead to performance degradation in extreme cases. Additionally, for domains requiring highly specialized knowledge, the system's performance may be limited as the LLM agent's knowledge base might not cover all domain-specific details. The user interface and interaction mechanisms of the system may require further optimization to accommodate different user habits and needs.