RODS Dynamically Synthesizes Training Data at the Agent's Capability Boundary for Multi-Turn RL

RODS addresses static dataset depletion in multi-turn tool-use RL by using rollout reward variance — derived from the Popoviciu bound in GRPO — as a zero-cost signal to identify and synthesize new tasks at the agent's current capability boundary. This keeps informative gradient signal flowing as the policy improves.