Reinforcement Learning from Collective Rewards (RLCR) is a new training framework that develops language models into collaborative agent organizations that share a workspace and message board. GenReasoning used the method to train the 120B parameter Avalon-M1-Mini to study how models interact within this collective structure.
The company reported that the training led to the emergence of organizational hierarchies, divisions of labor, and increased verification through the recruitment of specialized testers and verifiers. GenReasoning stated the RLCR framework is intended to act as a laboratory for studying collective intelligence and the safety of multi-agent systems.
