This AI paper introduces MAA (Multi-Agent Architecture Search): A new machine learning framework that optimizes multi-agent systems
Big Language Model (LLM) It is the basis of a multi-institutional system that allows multiple AI agents to collaborate, communicate and solve problems. These agents use LLM to understand tasks, generate responses and make decisions, mimicking human teamwork. However, efficiency lags when executing these types of systems because they are based on fixed designs that not all tasks change, resulting in them using too many resources to deal with simple and complex problems, wasting computations and resulting in slow responses . Therefore, this creates significant challenges in handling diverse tasks while trying to balance accuracy, speed and cost.
Currently, multi-institutional systems rely on existing methods camel, autogene, metagpt, dspy, evoprompting, gptswarm, and voagentfocusing on optimizing specific tasks such as timely adjustments, proxy analysis and communication. However, these approaches are difficult in adaptability. They follow a pre-fixed design without tuning various tasks, so they are somewhat inefficient in handling complex and simple queries. They lack flexibility through manual methods, while automation systems can only target the best configuration for search without dynamically adjusting efficiency. This makes these methods expensive in computing and leads to lower overall performance when applied to real-world challenges.
To solve the limitations of existing multi-agent systems, the researchers proposed MAA (Multi-Proxy Architecture Search). This framework uses probabilistic proxy supernet to generate query-related multi-proxy architectures. Rather than choosing the best system for fixedness, Maas A multi-agent system that dynamically sample customization for each query, balance performance and compute cost. The search space is defined by proxy operators, who are LLM-based workflows involving multiple agents, tools, and tips. SuperNet understands the distribution of possible proxy architectures and optimizes them based on task utilities and cost limitations. The architecture of the controller network sampling is used in querying, using a A mixture of Experts (MOE)– Effectively selected style mechanism. The framework is optimized through cost-attracting experience Bayesian Monte Carloupdate the proxy operator using a text gradient-based method. The framework provides automated multi-agent evolution that can improve efficiency and adaptability when dealing with diverse and complex queries.
Researchers evaluate Maas Six public benchmarks Cross-mathematical reasoning (GSM8K, Mathematics, Multiarith), Code generation (Humaneval, MBPP), and Tool usage (Gaia)and 14 baselinesincluding single agent method, handmade multi-agent system and automation method. Maas always outperform all baselines, achieving average optimal score 83.59% Cross-tasks and major improvements 18.38% exist Gaia Level 1 Task. Cost analysis shows that MAA’s resources are effective, requiring minimal training tokens, lowest API costs and shortest wall lock time. The case study highlights its adaptability when dynamically optimizing multi-agent workflows.
In summary, this method uses proxy supernets that adapt to different queries to fix problems in traditional multi-proxy systems. This makes the system work better, use resources wisely, and becomes more flexible and scalable. In future work Maas It can be developed into a flexible but extended framework to improve automation and self-organization in future work. Future work may also be optimized in sampling strategies, improving domain adaptability, and incorporating real-world constraints to improve collective intelligence.
Check Paper. All credits for this study are to the researchers on the project. Also, don’t forget to follow us twitter And join us Telegram Channel and LinkedIn GrOUP. Don’t forget to join us 75K+ ml reddit.
Recommended open source AI platform: ‘Intellagent is an open source multi-proxy framework that evaluates complex dialogue AI systems‘ (Promotion)
This AI paper introduces MAAS (Multi-Agent Architecture Search): a new machine learning framework that optimizes the first multi-agent system on Marktechpost.
