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

L1 core-mechanism diagnostic. Runs the four-signal gate. Exploratory results do not establish a stable win.

Paper information¶

Field	Details
Paper	A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients
Organization	First-author affiliation: Xiaohongshu
First public release	2026-09-10
Authors' code	No public author repository was found on 2026-09-14.
Adapter / method	`adaptive-opd-gate`
Local reproduction code	`src/auto_research/post_training/latest_20260914.py`

Original paper¶

Motivation and method¶

Combine entropy, uncertainty, reference deviation, and teacher–student gap in a per-token gate, adaptively assigning weight between forward and reverse KL.

flowchart LR
  I[Input and public state] --> M[Adaptive OPD Gate mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or update]

Key figure from the paper¶

Figure 1: Training flow and key optimization steps. From the original paper; copyright remains with the authors.

Core computation or operation¶

The local deterministic reference kernel implements the decisive operation and reports gate, selection, projection, graph-edge, or state-update statistics. Tests cover formula invariants and boundaries.

Paper results¶

The paper reports wins in 33/36 exploratory cells and 19/26 mean-matched static comparisons; some main three-seed comparisons were not significant. These paper results are not directly comparable with the scaled-down local run. No online gain is inferred without an online A/B test.

Local reproduction¶

See the three-seed metrics. All candidates use identical public fixtures, budgets, and seeds. These results establish only execution of the core state transitions.

Reproduction boundaries¶

The four-signal gate runs locally. Exploratory diagnostics are not stable performance gains.