Two Minute Papers · 2026-06-19 · notable
Two Minute Papers: 'Scientists Found A Better Language For AI Agents'
Two Minute Papers covers RecursiveMAS, a UIUC/Stanford/NVIDIA/MIT framework that lets multi-agent systems trade latent thoughts instead of text and reports 2.4x faster inference with 75.6% fewer tokens.
Two Minute Papers walks through RecursiveMAS, a multi-agent framework that swaps text messages for shared latent thoughts.
What is it?
A Two Minute Papers video from Karoly Zsolnai-Feher explaining RecursiveMAS, a multi-agent system from UIUC, Stanford, NVIDIA, and MIT. RecursiveMAS connects specialist LLM agents through a small RecursiveLink module so they exchange compressed latent states across reasoning rounds instead of long text turns.
How does it work?
RecursiveMAS treats the whole agent team as one recursive computation. Each agent reads and writes shared latent thoughts through RecursiveLink, supports four patterns (sequential, mixture, distillation, deliberation), and only trains about 0.31% of the system's parameters while the rest of the agents co-evolve.
Why does it matter?
On the reported benchmarks RecursiveMAS lifts accuracy by 8.3 points on average, runs 2.4x faster end to end, and cuts token use by 75.6% at the deepest recursion depth, directly attacking the cost and latency that have kept production multi-agent stacks small.
Who is it for?
Builders of multi-agent and tool-using LLM systems
Try it
https://github.com/RecursiveMAS/RecursiveMAS