Expert Personas Reduce LLM Factual Accuracy

Researchers at the University of Southern California publish a preprint reporting that persona-based prompting improves alignment but reduces factual accuracy on knowledge-heavy tasks. Using MMLU tests, persona prefixes cut multiple-choice accuracy to 68.0% versus 71.6% for the base model while boosting safety guardrails like JailbreakBench refusal rates by 17.7 percentage points. They propose PRISM, a gated LoRA routing method to balance trade-offs.