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.
Related event: Stanford's RecursiveMAS Recurses Multi-Agent Systems in Latent Space(2 posts)→
More from coding & agent
- Generic API-catalog MCP server vs one tool per job: a context-bloat experiment — MandeepKaur9407 · 2026-11-16
- Berkeley paper: LLMs answer with stale values even when the update is in context — ombre-ai · 2026-11-16
- Effective HTML: 4.2k-star agent skills turn wireframes into self-contained interactive prototypes — tom_doerr · 2026-11-16
- skills-gate: one gate skill fronts 148 agent skills at just 389 tokens — Distinct-Pie2389 · 2026-11-16
- Open-weight Karpo-Decision-31B tops Decision Index at 73.9, one forward pass per answer — CodeByPoonam · 2026-11-16
- shadcn boosts @uisdk: one-shot UI rebuild that's instantly E2E-testable by AI — shadcn · 2026-11-16