RLCR trains an LLM agent org with a CEO and specialists via one shared reward stream
rosstaylor90 · x · 2026-10-23
GenReasoning introduces RLCR (Reinforcement Learning from Collective Rewards), a recipe that trains a language model to act as an agent organization—sharing a message board and workspace—driven by a single collective reward stream. One policy plays both the CEO and the specialists it hires; they used it to train the 120B Avalon-M1-Mini.
Key findings:
- Emergent hierarchy and division of labor: the CEO learned to hire lead ICs, mirroring Brooks' surgical team model
- More collective verification: agents hired verifiers and testers
- Rise of "collective thinking": agents reason more about peers, offer help, and do more peer review
The team positions RLCR as a laboratory for studying collective intelligence and multi-agent safety; a blog post and technical report are available.
More from coding & agent
- Cursor agents can now drive your desktop, and one dev warns self-verifying agents just move the quality gate — petrusenko_max · 2026-10-24
- End of 'good enough software': formal verification becomes the missing AI demand curve — bookwormengr · 2026-10-24
- Dev says agents now run everything in his workflow except scrolling Shorts — tokenbender · 2026-10-24
- DAB, a cross-database agent benchmark, hits EMNLP Oral; best model only 38% pass@1 — adityagp · 2026-10-24
- Firecrawl Alexandria launches sports data APIs across 14 providers after $75M Series B — devdigest · 2026-10-24
- 27B open-source model beats GPT-5.5 and Opus 4.8 on desktop tasks at 1.5% of the cost — alex_verem · 2026-10-23