RLCR trains LLMs as agent organizations: emergent hierarchies, division of labor, verifiers
dair_ai · x · 2026-10-24
The GenReasoning team introduces RLCR (Reinforcement Learning from Collective Rewards), a recipe training language models as agent organizations. Avalon-M1-Mini (120B) runs multiple agents sharing a message board and workspace.
Key findings:
- 🛠️ Emergent organizational hierarchies and division of labor
- 🕵️ Increased collective verification via 'hires' of verifiers and testers
- 👥 Rise of collective thinking: agents reason about helping other agents
Benchmarks show the eight-hire organization beats Inkling Small (a larger model distilled from a stronger teacher) on 6 of 18 benchmarks and ties on 3, generally outperforming both the single-agent setup and Nemotron 3 Super on the same base. The team positions RLCR as a laboratory for studying collective intelligence and multi-agent safety. omarsar0 notes coordination should be a native agent capability rather than something bolted on via harnesses.
More from Models
- Anthropic secretly A/B tests Claude limits: same $200 plan, ~20% less usage — RexDouglass · 2026-10-24
- Kimi K3.1 Pro Benchmarks: Tops SWE Marathon at 51.7, Near Opus 5.5 Across Coding Boards — scaling01 · 2026-10-24
- Kimi 3.1 priced at $2/$10 per 1M tokens, still not competitive vs DeepSeek, says investor — bindureddy · 2026-10-24
- NVIDIA's Nemotron 3.5 Bolt lands on Ollama: small VLM with 128K context — ollama · 2026-10-24
- LlamaIndex's OCRia-1 matches top parsers at 1/4 the cost, 7x faster than Claude — llama_index · 2026-10-24
- Experts say OpenAI's redacted paper proof strategy is likely irreparable — teortaxesTex · 2026-10-24