From Imitation to Reward Discovery: On-Policy Warmup for Agentic RL
The findings support on-policy distillation as an effective warmup for agentic RLVR and identify initial reward discovery as a mechanism that can contribute to the observed acceleration.