SEP 10, 2026 · PREPRINT
A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients
arXiv
Early-phase methodological comparison of gating parameterizations in knowledge distillation with exploratory aggregate evidence from a single-seed sweep, targeted replications underpowered (n=3), and results presented as a coordinate system rather than definitive findings.
Reported
Configurations outperforming sing…33 of 36 comparable cells
Dynamic gating ahead of static ba…19 of 26 cells
Replication sample size for headl…3