Self-Activating Plug Layers: Teaching New Skills to a Frozen Vietnamese Decision Model Without Forgetting
Year: 2026
Type: preprint
Abstract: A frozen 22-layer ModernBERT typed-decision model (Laya-style, Vietnamese/English) receives a zero-initialised copy of layer 21 as a plug layer, gated per sentence by a linear detector on the layer-10 [CLS] vector. Pre-registered dev results with paired bootstrap: rumour rebuttal +21.4 / +19.3 points on two bodies; two new Vietnamese skills +20.0 (SREDFM) and +29.0 (VHE event typing) points; groups the detector does not fire on are unchanged (exactly zero on all 142 dev sets in the latest run; one non-significant exception in the first gated run). Negative results are included. The record contains the paper, mechanism code, per-item predictions, scoring files, logbook excerpts, a re-scoring script and checksums. License: paper under CC-BY 4.0; code under Apache-2.0.
Author: Ho, Hiep
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225792019
Cites:
Cited by:
Related to: