Unified framework compares token-level gating methods for AI model distillation

Original: A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients

Source: arxiv.org ↗