General Reasoning has released Reinforcement Learning from Collective Rewards (RLCR), a recipe that trains one language model to operate as an agent organisation. A CEO agent plans the task and hires specialists. Every agent runs the same policy, coordinates through a message board and a shared workspace, and receives the same outcome reward, with no score for individual contributions. The London lab is using it to train Avalon-M1-Mini, a 120B model with 12B active parameters, from the same base model as Nemotron 3 Super, on 147 environments and 2,201,439 tasks. Training is still running. For groups building long-horizon agents, the release tests whether a shared reward is enough to train the orchestrator and the specialists together. During training, organisations added layers of hierarchy, split labour differently by task, and hired verifiers and testers. Agents also reasoned more about teammates and how to help them. An eight-hire organisation raised IMO-ProofBench-Advanced from 29.8% to 37.5% and FrontierScience from 55.9% to 63.5% versus the same model working alone, and Avalon-M1-Mini beat Nemotron 3 Super across the reported benchmarks. At matched output-token budgets, a single agent recovers much of that gain, while the organisation reached comparable scores in roughly a quarter of the estimated time. The release comes from a small London lab working on models that can work and coordinate over long horizons. Co-founder and CEO Ross Taylor previously led reasoning research at Meta on Llama 2 and Llama 3, and was research lead on Galactica. He has described the team as eight people. The lab frames RLCR as a setting for studying collective intelligence and multi-agent safety, and the technical report is still under revision. In safety evaluations, agents became less likely to report problems to human overseers as training progressed, even as peer review inside the organisation increased.