LLM-based multi-agent system is characterized by scheme, logic, equipment usage and memory capabilities, which form the foundation of applications such as chatbots, code generation, mathematics and robotics. However, these systems face significant challenges as they are manually designed, leading to high human resource costs and limited scalability. Graph-based methods have attempted to automate workflow designs by preparing the workflows as a network, but their structural complexity prohibits scalability. The state-of-the-art approaches represent multi-agent systems as programming code and use advanced LLM as meta-agents to customize workflows, but focus on working-level solutions that generate single working systems. This one size-fit-all approaches lack automatic adaptation ability to individual user questions.
LLM-based multi-agent systems are the foundations of various real-world applications, including code intelligence, computer use and intensive research. These systems have LLM-based agents equipped with planning capabilities, database access and tool function calls that collaborate to achieve promising performance. The initial approach focuses on adapting the signals or hyperpieters through the development algorithm to automate agents profiling. ADAS introduced the code representation for agents and workflows with meta-agents to generate workflows. In addition, Openai has developed advanced arguments in LLMS by developing the O1 model. Models such as QWQ, QVQ, Deepseek and Kimi have followed the suit developing architecture like O1. The O3 model of Openai achieves promising results on the Arg -gi benchmark.
The CAI Lab, Singapore, Chinese Academy of Sciences, National University of Singapore have been proposed by researchers and Shanghai Ziao Tong University, which has been designed to automate the construction of Query-Level Multi-Against System to Queri-Level Meta-class-agent, which manufactures a customized system according to the user Querry. Researchers disturbed the Deepsek R1 to supply the floworenner with the fundamental arguments capabilities required to create a multi-agent system, and then enhanced it through learning reinforcement with external execution reaction. A multi -purpose reward mechanism has been developed to optimize training in three important dimensions: performance, complexity and efficiency. This enables the flowers’ to generate individual multi-agent systems through deliberate arguments for each unique user querry.
Researchers select three datasets: for engineering-oriented tasks for detailed evaluation in various code generation scenarios for bigcodebench, humaneval, and MBPP algorithm challenges. Flowreasoner is evaluated against three categories of baseline:
- Single-model direct call using standalone LLM
- Manually designed workflows, including Self-Rifine, LLM-Rish and LLM-Blender, which have man-made arguments strategies
- Automatic workflow adaptation methods such as Aflow, Adas and Maas that create workflows through search or adaptation.
Both O1-Mini and GPT-4O-Mini are used as a worker model for manually designed workflows. Flowreasoner is applied as a worker model using O1-Mini with two variants of Leepseek-R1-DISTILL -QWen (7B and 14B parameters).
Flowreasoner-14B improves all competitive approaches, the strongest baseline, achieves a overall improvement of 5 percent marks compared to MAA. This is more than the performance of its underlying worker model, O1-Min, by a sufficient difference of 10%. These results reflect the effectiveness of the workflow-based logic structure in increasing code generation accuracy. To evaluate generalization capabilities, the experiments are replaced by the O1-MINI worker with models such as Qwen2.5-Coder, Cloud, and GPT-4O-Min, while the meta-agent is fixed as either Floresner-7B or Floidrasenar-14B. Floresenar displays remarkable transferable, maintaining continuous performance in various activists models on the same tasks.
In this letter, researchers presented Floresner, a query-tier meta-agent designed to automate the construction of individual multi-agent systems for individual user questions. Flowreasoner focuses on performance, complexity, and efficiency using external execution reaction and reinforcement learning, which focuses on complicated search algorithms or carefully designed search sets to generate customized workflows without relying on the search sets. This approach reduces human resource costs, while enabling more adaptive and efficient multi-agent systems enhances scalability that dynamically optimize their structure depending on specific user questions rather than relying on certain workflows for the entire work categories.
Check it Paper and Github pageAlso, don’t forget to follow us Twitter And join us Wire And LinkedIn GROUPDon’t forget to join us 90K+ ML Subredit,
[Register Now] Minicon Virtual Conference on Agentic AI: Free Registration + attendance Certificate + 4 hours small incident (May 21, 9 am- 1 am PST)
Sajjad Ansari is a final year graduation from IIT Kharagpur. As a technical enthusiast, he delays practical applications of AI with focus on understanding the impact of AI technologies and their real -world implications. He aims to clarify complex AI concepts in a clear and accessible way.
