Unified Per-Token Gating for On-Policy Distillation

A new four-coefficient parameterization for per-token gating in on-policy distillation combines multi-channel composition with explicit bias. When distilling a Qwen3-32B teacher to a Qwen3-4B student on TweetEval, this method outperformed standard 1D entropy or gap-only gating in 33 of 36 test cells.