FlowSearch: Advancing deep research with dynamic structured knowledge flow

TL;DR

FlowSearch tackles open-ended scientific research by modeling reasoning as a dynamic knowledge flow over a directed acyclic graph, with subtask nodes and dependency edges. It couples a Knowledge Flow Planner, Knowledge Collector, and Knowledge Flow Refiner to iteratively expand, execute, and reconfigure the flow, enabling parallel exploration and hierarchical task decomposition. Empirically, FlowSearch achieves state-of-the-art results on GAIA, HLE, and TRQA and competitive performance on GPQA, driven by graph-guided planning and tool-enabled reasoning. This approach advances autonomous, reflective scientific discovery with scalable, auditable workflows across disciplines.

Abstract

Deep research is an inherently challenging task that demands both breadth and depth of thinking. It involves navigating diverse knowledge spaces and reasoning over complex, multi-step dependencies, which presents substantial challenges for agentic systems. To address this, we propose FlowSearch, a multi-agent framework that actively constructs and evolves a dynamic structured knowledge flow to drive subtask execution and reasoning. FlowSearch is capable of strategically planning and expanding the knowledge flow to enable parallel exploration and hierarchical task decomposition, while also adjusting the knowledge flow in real time based on feedback from intermediate reasoning outcomes and insights. FlowSearch achieves state-of-the-art performance on both general and scientific benchmarks, including GAIA, HLE, GPQA and TRQA, demonstrating its effectiveness in multi-disciplinary research scenarios and its potential to advance scientific discovery. The code is available at https://github.com/Alpha-Innovator/InternAgent.

Figures (4)

• Figure 1: FlowSearch (Ours) achieves leading performance on the GAIA, GPQA, and HLE benchmarks, outperforming competitive agent workflow methods (OpenAI-DeepResearch, MiroFlow, Manus, OWL) as well as LLM-based approaches (GPT-5, Intern-S1, DeepSeek-R1).
• Figure 2: Overview of FlowSearch. Top part illustrates the Knowledge Flow Planning process, where the Knowledge Flow Planner incrementally expands the structured knowledge flow. Middle part depicts the iterative process of Knowledge Collection and Flow Refinement, where nodes are executed by the Knowledge Collector and the flow is dynamically adjusted by the Knowledge Flow Refiner based on newly acquired knowledge. Lower part highlights the three key components of FlowSearch—Flow Planner (left), Flow Refiner (center), and Knowledge Collector (right)—and their collaborative role in enabling systematic, adaptive, and efficient deep research.
• Figure 3: Performance on the TRQA benchmark. FlowSearch (Ours) significantly outperforms previous works.
• Figure 4: Case study comparing the conventional deep research framework OWL with our FlowSearch on a scientific question.