Persona-Level Safety in Abliterated LLMs: Can Declarative Identity Anchors Defend When Model Guardrails Are Gone?

Read full text Summarise with VIRAT
Authors: Tom Lee, Jihong Lee
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

First empirical study of Declarative Identity Anchors (structured persona files) as a safety mechanism in abliterated LLMs. Using a 2×2 factorial design with Qwen 3.5 9B, we find persona constraints improve aligned model safety (+33pp, 50%→83%) but provide minimal protection in abliterated models (+6pp). We identify the Helpful Assistant Paradox and category-specific effectiveness patterns. v3: Gemini review feedback — tone adjustments, LLM-as-Judge limitation, Helpful Assistant Paradox quantitative support, Appendix truncation clarification, updated affiliation.

Fields of Study & Subjects

Ethics and Social Impacts of AI Adversarial Robustness in Machine Learning Topic Modeling

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
21 March 2026
Language
EN
Document type
Preprint