Collective training

General Reasoning introduces training recipe for agent organizations

GenReasoning introduced Reinforcement Learning from Collective Rewards, a recipe for training language models as agent organizations that share a message board and workspace. The account said it used the approach to train Avalon-M1-Mini, a 120B model.

GenReasoning reported learned hierarchies, divisions of labor, more collective verification and increased attention to teammates' needs. rosstaylor90 described the findings as early work from a team of 8 people, with further experiments, ablations and training underway.