One Sigmoid, Smarter Learning: Implementing SDAR's Asymmetric Knowledge Gate

SDAR Gate solves a critical problem in reinforcement learning: selectively distilling knowledge from a privileged teacher without amplifying noise. Using a single sigmoid function to measure teacher-student confidence gaps, the mechanism enables dense token-level guidance that significantly improves agent learning efficiency. This asymmetric approach is particularly valuable for multi-step agents where traditional reward signals are too coarse-grained.