Prompt-based persona safety rankings do not predict activation-steering vulnerabilities across LLM architectures, exposing a prosocial persona paradox and the incompleteness of single-method evaluations.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMs
Prompt-based persona safety rankings do not predict activation-steering vulnerabilities across LLM architectures, exposing a prosocial persona paradox and the incompleteness of single-method evaluations.