S3 proposes cooperative reinforcement learning to train agent swarms

S3 researchers propose training multi-agent swarms with cooperative reinforcement learning, so agents learn when to work alone, share findings, or read teammates' work instead of following a plan written by developers.

Elvis Saravia reports the claims on X: with matched budgets, S3 swarms tested at twice their training size solve over 25 times more coordination problems than the strongest baseline. In parallel exploration, communicating agents repeat nearly 50% less work and beat best-of-N by up to 9.4 points.

These figures come from the S3 proposal as relayed in a single X post; the post does not give the benchmarks or baselines behind them.