Multi-agent systems based on LLM characterized by planning, reasoning, the use of tools and memory capacities form the basics of application such as chatbots, code generation, mathematics and robotics. However, these systems are faced with significant challenges because they are manually designed, leading to high costs of human resources and a limited scalability. The methods based on graphics have attempted to automate workflow conceptions by formulating workflows as networks, but their structural complexity restricts scalability. Railing approaches represent multi-agent systems such as programming code and use advanced LLMs as meta-agents to optimize workflows, but focus on solutions at the tasks that generate specific tasks systems. This unique approach does not have the automatic adaptability to individual user requests.
LLM-based multi-agent systems are the basis of various real world applications, including code intelligence, computer use and in-depth research. These systems include agents based on LLM equipped with planning capacities, access to the database and the invocation of the tool function that collaborates to obtain promising performance. The first approaches focused on optimizing prompts or hyperparameters through evolution algorithms to automate the profiling of agents. Adas introduced the representation of the code for agents and workflows with a meta-agent to generate workflows. In addition, Openai has advanced reasoning in LLM by developing the O1 model. Models like QWQ, QVQ, Deepseek and Kimi have followed suit, developing O1 -type reasoning architectures. OPENAI's O3 model obtains promising results on the ARG-AGI reference.
Researchers from the SEA AI Lab, Singapore, the University of the Chinese Academy of Sciences, the National University of Singapore and the Shanghai Jiao Tong University proposed Flowreasoner, a request meta-agent designed to automate the creation of multi-agent level of request systems, generating a personalized system by question. The researchers distilled Deepseek R1 to provide Flowreason with the fundamental reasoning capacities necessary to create multi-agent systems, then improve it by learning to strengthen external execution. A versatile reward mechanism is developed to optimize training through three critical dimensions: performance, complexity and efficiency. This allows Flowreasoner to generate personalized multi-agent systems thanks to deliberative reasoning for each single user query.
Researchers select three data sets: Bigcodebench for engineering -oriented tasks, Humaneval and MBPP for algorithmic challenges for a detailed assessment in various code generation scenarios. Flowreasoner is evaluated in three categories of basic lines:
- Direct invocation with unique model using autonomous LLMS
- Workflows manually designed, including self-refine, LLM-Debate and LLM-Blender with reasoning strategies designed by humans
- Automated workflow optimization methods like Aflow, ADAS and MAAS which build work flows through research or optimization.
O1-Mini and GPT-4O-MINI are used as workers models for manually designed workflows. Flowreasoner is implemented with two variants of depth-R1-Distill-Qwen (parameters 7b and 14b) using O1-mini as a worker model.
Flowreason-16b surpasses all competing approaches, making a global improvement of 5 percentage points compared to the strongest base line, Maas. It exceeds the performance of its underlying worker model, O1-Mini, by a substantial margin of 10%. These results show the effectiveness of the reasoning frame based on the workflow to improve the accuracy of the code generation. To assess the generalization capacities, experiences are carried out to replace the worker O1-Mini with models such as QWEN2.5-Coder, Claude and GPT-4-MINI, while keeping the fixed meta-agent like Flowreason-7B or Flowreasoner-14b. Flowreasoner has significant transferability, maintaining coherent performance on different models of workers on the same tasks.
In this article, researchers present Flowreasoner, a meta-agent at the level of the query designed to automate the creation of personalized multi-agent systems for individual user requests. Flowreasoner uses external feedback and learning to strengthen versatile rewards by focusing on performance, complexity and efficiency to generate optimized workflows without relying on complex research algorithms or carefully designed research complexes. This approach reduces human resources costs while improving scalability by allowing more adaptive and effective multi-agent systems that dynamically optimize their structure according to specific user requests rather than relying on fixed workflows for whole tasks.
Discover the Paper And GitHub page. Also, don't forget to follow us Twitter And join our Telegram And Linkedin Group. Don't forget to join our 90K + ML Subdreddit.
Sajjad Ansari is a last year's first year of the Kharagpur Iit. As a technology enthusiast, he plunges into AI's practical applications by emphasizing the understanding of the impact of AI technologies and their real implications. It aims to articulate complex AI concepts in a clear and accessible way.
