s³: cooperative multi-agent RL lets swarms self-organize, beating baselines 25x
omarsar0 · x · 2026-11-09
A University of Washington / Allen AI paper introduces s³, the first multi-agent RL method where agents learn when to work alone, share findings, or read teammates' work via cooperative RL with no predefined roles. At matched budgets, swarms at 2x training size solve over 25x more coordination problems than the strongest baseline; communicating agents cut duplicated work by 50% and beat best-of-N by up to 9.4 points.
More from coding & agent
- Solo dev's infinitely evolving Claude-built Backrooms game hits 35K players in 36 hours — minchoi · 2026-11-09
- RentAHuman lets AI agents hire real people for physical-world tasks via MCP API — conitzer · 2026-11-09
- Jane Street open-sources /socratic-tutor: an agent skill that teaches via questioning, not lecturing — daniel_mac8 · 2026-11-09
- Fly-by-Code: coding agents pilot drones via active visual and physical feedback loops — jbhuang0604 · 2026-11-09
- Unreal Engine 5.9 preview is the first UE release built primarily with agentic workflows — nicolascraske · 2026-11-09
- YC-backed Floworks launches Eniac, an API for verified large-scale web data extraction — ycombinator · 2026-11-09