TLDR: FlowSearch is a multi-agent AI framework that uses a dynamic, structured knowledge flow to conduct deep research. It plans tasks, collects information, and refines its strategy in real-time, enabling parallel exploration and hierarchical task decomposition. This approach allows it to achieve state-of-the-art performance on complex general and scientific benchmarks by effectively managing knowledge and adapting to new findings.
Deep research, whether in general problem-solving or scientific discovery, is a complex endeavor. It demands both broad exploration of knowledge and deep, multi-step reasoning. Traditional AI systems often struggle with these challenges, facing issues like “tunnel vision” in single-agent setups or context overload in multi-agent systems with linear execution.
Introducing FlowSearch: A New Approach to AI-Driven Research
A team of researchers from the Shanghai Artificial Intelligence Laboratory has introduced FlowSearch, a novel multi-agent framework designed to tackle the complexities of deep research. FlowSearch stands out by actively building and evolving a “dynamic structured knowledge flow” to guide its research process. This innovative approach allows AI agents to strategically plan, explore in parallel, break down tasks hierarchically, and adapt their strategies in real-time based on new information.
How FlowSearch Operates
FlowSearch is built around three core components that work together in an iterative loop:
The Knowledge Flow Planner
At the outset, the Knowledge Flow Planner creates an initial “knowledge flow.” Think of this as a dynamic roadmap where each “node” represents a subproblem to solve or a key concept to retrieve, and “edges” show the dependencies between them. This isn’t a static plan; it can be expanded incrementally as the research progresses, ensuring both broad exploration and deep investigation.
The Knowledge Collector
Once the plan is in place, the Knowledge Collector identifies subtasks that are ready to be executed (meaning all their prerequisites are met). These subtasks are then assigned to specialized “executor agents.” These agents, powered by large language models and equipped with various tools like web browsing, document extraction, and even visual question answering, work to resolve their assigned nodes. If successful, the gathered or derived knowledge is summarized and added to the node’s context.
The Knowledge Flow Refiner
After knowledge is collected, the Knowledge Flow Refiner steps in. This component acts as a reflective mechanism, analyzing the current state of the knowledge flow and making dynamic adjustments. It can add new tasks, remove redundant ones, modify existing subtasks, or adjust dependencies between them. This real-time adaptation ensures the research stays on track, efficient, and responsive to new insights.
Why FlowSearch Excels
Unlike conventional linear research pipelines that can get overwhelmed by too much information or miss crucial connections, FlowSearch’s graph-based approach explicitly models task dependencies. This allows for parallel execution of independent branches, more efficient knowledge propagation, and the ability to maintain global coherence while performing deep reasoning on local subproblems. The system’s ability to dynamically refine its plan based on intermediate findings makes it highly adaptive and robust for open-ended research.
Also Read:
- FlashResearch: Accelerating AI-Powered Deep Research Through Dynamic Orchestration
- DeepEvolve: An AI Agent That Learns and Builds Better Algorithms
Demonstrated Impact
FlowSearch has demonstrated impressive capabilities across a range of challenging benchmarks. It has achieved state-of-the-art performance on general problem-solving tasks like GAIA, and excelled in multi-disciplinary scientific question-answering benchmarks such as HLE, GPQA, and TRQA. These results highlight its effectiveness in diverse research scenarios and its potential to significantly advance scientific discovery. For instance, FlowSearch (o4-mini) achieved a GAIA score of 76.96%, significantly outperforming the same model without its agentic framework (16.97%), underscoring the importance of structured task decomposition and flow-based execution over just model size.
The full details of this innovative system can be found in their research paper: FlowSearch: Advancing deep research with dynamic structured knowledge flow.
The authors of this groundbreaking work are Yusong Hu, Runmin Ma, Yue Fan, Jinxin Shi, Zongsheng Cao, Yuhao Zhou, Jiakang Yuan, Xiangchao Yan, Wenlong Zhang, Lei Bai and Bo Zhang.
