RecursiveMAS: Agent swarms share latent thoughts like looped Transformers, cutting tokens by 75%
james_y_zou · x · 2026-09-30
A NeurIPS 2026 paper from James Zou's team, RecursiveMAS, extends the recurrent-depth idea of looped Transformers from a single model to multi-agent systems:
- A team of heterogeneous LLM agents collaborates like a looped Transformer: each agent acts as a layer, passing and refining latent states across collaboration rounds via a lightweight RecursiveLink module, instead of communicating only through text.
- An inner–outer loop training algorithm teaches the team to collaborate as a whole through recursion.
- Results: +8.3% average accuracy across all benchmarks; 2.4× end-to-end speedup and −75.6% token usage vs. text-based MAS; 5 collaboration styles across 9 benchmarks.
- Authors span Stanford, UIUC, NVIDIA, and MIT; code, models, data, an interactive playground, and a tutorial video are released.
- The goal: make recurrent collaboration a new scaling axis for agent systems.
More from coding & agent
- Developer Has Dots Drive Its Own Computer to Run Blender and Model Itself in 3D — Dimillian · 2026-09-30
- GPT-Sol 6.1 Ships Too, as Omarsar Argues Codex-Dots Combo Unlocks New Agent Workflows — omarsar0 · 2026-09-30
- LangSmith adds Trajectories: collapsing agent sessions into readable, ordered execution paths — LangChain · 2026-09-30
- Codex Cloud announced at OpenAI DevDay, available across Plus through Enterprise tiers — testingcatalog · 2026-09-30
- Dots Roll Out as Slack-Embedded Specialists That Take On Full Jobs Inside Companies — danshipper · 2026-09-30
- Developer: Always-On AI Agents Like Muse and Grok Bot Must Be Open Source — yacineMTB · 2026-09-30