#RL initialization

'Early Experience uses agent-generated future states as supervision to train language agents without rewards or large demo sets, outperforming imitation learning across eight benchmarks and improving RL initialization.'