GenReasoning introduces RLCR, training language models as agent organisations with shared message boards

GenReasoning has introduced Reinforcement Learning from Collective Rewards (RLCR), a method that trains language models as agent organisations sharing a message board and workspace. The team used RLCR to train Avalon-M1-Mini, a 120B-parameter model, and studied its behaviour. Reported findings include learned organisational hierarchies and divisions of labour, increased collective verification through the emergence of verifier and tester roles, and a rise in agents reasoning about other agents and how to assist them. The authors say RLCR could serve as a laboratory for studying collective intelligence and multi-agent safety. The work remains under revision as training continues.

✨ AI summary of the linked source — not editorial content.

Explore the topic

• AI Agents