rl-optimization

SLCA-GRPO introduces Segment-Locked Credit Assignment to improve reinforcement learning for tool-calling agents by decoupling advantage estimation and using hierarchical rewards, leading to faster convergence and higher accuracy.