Expert Personas Improve LLM Alignment but Damage Accuracy: Bootstrapping Intent-Based Persona Routing with PRISM

TL;DR

This paper investigates how expert personas influence LLM performance, revealing that while they improve alignment and safety, they can reduce accuracy, and introduces PRISM, a method to optimize persona routing without external data.

Contribution

The paper provides a comprehensive analysis of expert persona effectiveness and introduces PRISM, a novel self-distillation pipeline for intent-based persona routing in LLMs.

Findings

Expert personas improve alignment and safety.

Expert personas can damage accuracy in some tasks.

PRISM effectively leverages expert personas without external data.

Abstract

Persona prompting can steer LLM generation towards a domain-specific tone and pattern. This behavior enables use cases in multi-agent systems where diverse interactions are crucial and human-centered tasks require high-level human alignment. Prior works provide mixed opinions on their utility: some report performance gains when using expert personas for certain domains and their contribution to data diversity in synthetic data creation, while others find near-zero or negative impact on general utility. To fully leverage the benefits of the LLM persona and avoid its harmfulness, a more comprehensive investigation of the mechanism is crucial. In this work, we study how model optimization, task type, prompt length, and placement can impact expert persona effectiveness across instruction-tuned and reasoning LLMs, and provide insight into conditions under which expert personas fail and succeed. Based on our findings, we developed a pipeline to fully leverage the benefits of an expert persona, named PRISM (Persona Routing via Intent-based Self-Modeling), which self-distills an intent-conditioned expert persona into a gated LoRA adapter through a bootstrapping process that requires no external data, models, or knowledge. PRISM enhances human preference and safety alignment on generative tasks while maintaining accuracy on discriminative tasks across all models, with minimal memory and computing overhead.