Jiaru Zou’s Post

🔥 Introducing 𝐑𝐞𝐜𝐮𝐫𝐬𝐢𝐯𝐞𝐌𝐀𝐒 — a new learning framework for scaling multi-agent system collaboration through latent-space recursion. 🌐 Website: https://lnkd.in/gixdr4hs 📄 Paper: https://lnkd.in/gjdnDuma 💻 Code: https://lnkd.in/gtxSSeEz 💡𝑪𝒐𝒓𝒆 𝑸𝒖𝒆𝒔𝒕𝒊𝒐𝒏: Can agent collaboration itself be scaled through recursion? RecursiveMAS answers this by treating the entire multi-agent system as a recursive latent computation, where agents iteratively refine shared reasoning states instead of merely exchanging text. ✨ 𝐊𝐞𝐲 𝐅𝐞𝐚𝐭𝐮𝐫𝐞𝐬 • 🔁 System-level latent recursion for multi-agent collaboration • 🔗 Lightweight RecursiveLink modules for connecting heterogeneous agents • ⚙️ Inner–outer loop training for whole-system co-optimization • 🧩 Plug-and-play across common MAS patterns • ⚡ Faster and cheaper inference via latent-space collaboration 📈 𝐄𝐦𝐩𝐢𝐫𝐢𝐜𝐚𝐥 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧𝐬 We evaluate RecursiveMAS across 9 benchmarks covering math, science, medicine, search, and code generation, and compare with strong single-agent, recursive, and multi-agent baselines. • 🏆 +8.3% average accuracy improvement • 🚀 1.2×–2.4× end-to-end inference speedup • 🪙 34.6%–75.6% token usage reduction Try it yourself! 👇 💻 Code: https://lnkd.in/gtxSSeEz 🤗 HF Models & Collections: https://lnkd.in/gyK23AuZ Please kindly give us a star ⭐ if you find RecursiveMAS useful! 💫 Huge thanks to all amazing collaborators from Stanford, UIUC, and MIT! James Zou , Jingrui He , Markus J. Buehler , Hanghang Tong , Tong Zhang , Pan Lu , Shizhe Diao , Jindong Jiang , Ruizhong Qiu , Rui Pan , Xiyuan Yang #RecursiveMAS #RecurisveLearning #MultiAgentSystems #ModelCollaboration #LLM #AgenticAI