How It Works
RL for Population Scaling
Training or coordinating many AI agents together as a group — a "population" or "swarm" — using reinforcement learning, instead of training and running a single agent alone. Researchers study how capability changes as the number of agents in the population grows, distinct from scaling a single model's parameters or training compute.
Origin · no single documented coiner
Builds on academic multi-agent reinforcement learning research, including DeepMind's 2019 paper "Malthusian Reinforcement Learning," which linked a sub-population's average return to its population size, and later work such as "Evolutionary Population Curriculum for Scaling Multi-Agent Reinforcement Learning" (ICLR 2020). "RL for Population Scaling" itself is not a standardized, widely cited term with one documented coiner; it is used informally to describe this area of research and the related discussion of how many-agent "swarms" of AI systems scale in capability.