General Reasoning releases RLCR: RL trains LMs as agent organisations with collective rewards
rosstaylor90 · x · 2026-10-23
General Reasoning introduced RLCR (Reinforcement Learning from Collective Rewards) and Avalon-M1-Mini (120B, fine-tuned from Nvidia Nemotron Super), where agents share a message board and workspace while the whole team shares one reward signal—no per-agent credit assignment.
Key findings
- Emergent organisational hierarchies and divisions of labour
- Increased collective verification, including 'hiring' verifiers and testers
- 'Collective thinking': agents reason more about helping other agents
- Shared rewards alone suffice: downstream performance keeps improving over long multi-agent rollouts
The team frames agent collectives as a way to tackle long-horizon tasks with lower latency, but also an extreme superintelligence technology with serious safety risks (citing this summer's OpenAI/HuggingFace incident), positioning RLCR as a laboratory for collective intelligence and multi-agent safety research.
More from coding & agent
- Figma MCP Server converts Figma designs into pixel-perfect React code — modelcontextprotocol · 2026-10-23
- Gemini 4 Argon builds a 3D Gothic cathedral from one image via Blender MCP — GoogleDeepMind · 2026-10-23
- Reddit thread asks: which tasks do your agents fail at, forcing you to pay humans? — Aggressive-Buyer8583 · 2026-10-23
- claim MCP server string-matches citations against sources and flags fabricated quotes — BondableColt18 · 2026-10-23
- AI agents take macOS kernel vulns to root: an XNU exploit dev case study — moyix · 2026-10-23
- Wavelength SDK hits mainnet, letting humans and AI agents send bitcoin with a few API calls — RSync25 · 2026-10-23