Popoviciu

Researchers propose RODS (Reward-driven Online Data Synthesis) to address the depletion of informative samples in multi-turn tool-use reinforcement learning. RODS uses progress reward variance as a bo…