Multi-agent systems often disappoint? The problem may not lie in the agents themselves, but in how they are organized. Most systems use fixed chain, tree, or graph structures that cannot be adjusted as tasks change. But what if the system could learn its own collaboration patterns? This groundbreaking paper introduces a framework called Puppeteer, which can dynamically orchestrate agents instead of relying on pre-designed topologies. The key point is: • Instead of predefined collaborative structures, a coordinator selects the next agent to speak based on the constantly changing dialogue state. • The REINFORCE algorithm training strategy directly optimizes the task success rate. • Instead of searching complex graph topologies, serialize all content into continuous agent choices, thus avoiding combinatorial complexity. The result was surprising: This naturally leads to a compact, cyclical pattern, rather than a large graph structure, in which 2-3 agents handle most of the work. Even more impressively, the system can autonomously detect efficiency issues. Results Showcase: • On GSM-Hard math problems: accuracy reaches 70% (compared to only 13.5% when the base model is used alone). • On MMLU-Pro: 83% (baseline 76%). • In SRDD software development: 76.4% (baseline 60.6%). These improvements are accompanied by a reduction in token costs. The paper shows that the token cost continues to decrease throughout the training process, while the performance improves. They also proved that the agent selection process satisfies the Markov property, which means that the current state can determine the optimal next agent without having to track the entire history. so: For AI developers, simplicity in learning is better than elaborate complexity. A trained router, along with some specialized intelligent agents, can outperform a well-designed workflow while reducing computational overhead.
This content was gexaicreator.com https://t.co/Gxsobg3hEN
Paperopenreview.net/pdf/9727f658d7…CdCeZbC
Thank you for taking the time to read this tweet! Follow @Yangyixxxx for AI information, business insights, and growth strategies. If you enjoyed this content, please like and share the first tweet to spread valuable information to more people.