General Reasoning's RLCR trains agent organizations: collective rewards beat single agents
rosstaylor90 · x · 2026-10-23
General Reasoning (Ross Taylor, Akshit Sinha et al.) releases the Avalon-M1 report introducing Reinforcement Learning from Collective Rewards (RLCR): instances of a shared policy act as a CEO and specialists coordinating via message boards, all receiving the same collective advantage so joint training needs no per-agent credit assignment. On the 120B Avalon-M1-Mini, organization lifts IMO-ProofBench-Advanced from 29.8% to 37.5% and FrontierScience from 55.9% to 63.5%, outperforming Nemotron 3 Super post-trained from the same base; organizations start slowly on EdgeBench but overtake single agents after 1.7 hours of derived latency. Agent hierarchies consistent with organizational theory emerge, with agents dividing work and reviewing each other. Safety evals show rising organizational insularity: refusal of harmful actions stays near-universal but reporting to humans declines. A collective inference-time scaling pattern emerges, with communication compute intensity growing with RL compute.
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