Representational Quarantine and Semantic Vaccination (RQSV)
Year: 2026
Type: article
Abstract: Representational Quarantine and Semantic Vaccination (RQSV) is a controlled causal research framework designed to evaluate whether weight-level safety interventions applied to a parent large language model influence the recurrence of specified safety failures in descendant models produced through standardized adaptation procedures. The framework reframes safety alignment as a lineage-level causal problem, treating the parent model as the site of intervention and independently adapted descendants as the units of outcome assessment. A structured literature audit of eighteen related works positions RQSV within existing research on safety degradation during fine-tuning, safety restoration, model merging, and activation steering while arguing that its randomized parent-to-descendant experimental design occupies a distinct methodological niche. The document specifies a three-branch experimental architecture comprising treated, sham-treated, and reference conditions; defines behavioral safety markers; outlines evaluation protocols and pilot-study requirements; and establishes explicit go/no-go criteria for empirical deployment. The work is presented as a research proposal and methodological framework intended for feasibility testing and preregistration prior to confirmatory experimentation.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225633808
FWCI:
Cites:
Cited by:
Related to: