Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Persona Switch: Mixing Distinct Perspectives in Decoding Time

Junseok Kim, Nakyeong Yang, Kyomin Jung

TL;DR

This work tackles the inconsistent gains of role-play prompting in LLMs by introducing Persona Switch, a decoding-time method that dynamically alternates between zero-shot and role-play outputs at each step using a confidence signal based on the average token-level top-two probability gap. The approach generates two candidate outputs per step and selects the more reliable one, enabling stepwise integration of distinct prompting perspectives without additional training. Across five reasoning benchmarks and multiple model families, Persona Switch achieves up to 5.13% higher accuracy than strong baselines and demonstrates that output confidence is a robust criterion for selecting reliable answers. The method is simple to implement, leverages existing prompts, and offers a practical framework for combining complementary prompting strategies to improve multi-domain reasoning in LLMs.

Abstract

Role-play prompting is known to steer the behavior of language models by injecting a persona into the prompt, improving their zero-shot reasoning capabilities. However, such improvements are inconsistent across different tasks or instances. This inconsistency suggests that zero-shot and role-play prompting may offer complementary strengths rather than one being universally superior. Building on this insight, we propose Persona Switch, a novel decoding method that dynamically combines the benefits of both prompting strategies. Our method proceeds step-by-step, selecting the better output between zero-shot and role-play prompting at each step by comparing their output confidence, as measured by the logit gap. Experiments with widely-used LLMs demonstrate that Persona Switch consistently outperforms competitive baselines, achieving up to 5.13% accuracy improvement. Furthermore, we show that output confidence serves as an informative measure for selecting the more reliable output.

Persona Switch: Mixing Distinct Perspectives in Decoding Time

TL;DR

This work tackles the inconsistent gains of role-play prompting in LLMs by introducing Persona Switch, a decoding-time method that dynamically alternates between zero-shot and role-play outputs at each step using a confidence signal based on the average token-level top-two probability gap. The approach generates two candidate outputs per step and selects the more reliable one, enabling stepwise integration of distinct prompting perspectives without additional training. Across five reasoning benchmarks and multiple model families, Persona Switch achieves up to 5.13% higher accuracy than strong baselines and demonstrates that output confidence is a robust criterion for selecting reliable answers. The method is simple to implement, leverages existing prompts, and offers a practical framework for combining complementary prompting strategies to improve multi-domain reasoning in LLMs.

Abstract

Role-play prompting is known to steer the behavior of language models by injecting a persona into the prompt, improving their zero-shot reasoning capabilities. However, such improvements are inconsistent across different tasks or instances. This inconsistency suggests that zero-shot and role-play prompting may offer complementary strengths rather than one being universally superior. Building on this insight, we propose Persona Switch, a novel decoding method that dynamically combines the benefits of both prompting strategies. Our method proceeds step-by-step, selecting the better output between zero-shot and role-play prompting at each step by comparing their output confidence, as measured by the logit gap. Experiments with widely-used LLMs demonstrate that Persona Switch consistently outperforms competitive baselines, achieving up to 5.13% accuracy improvement. Furthermore, we show that output confidence serves as an informative measure for selecting the more reliable output.
Paper Structure (39 sections, 4 equations, 1 figure, 12 tables)

This paper contains 39 sections, 4 equations, 1 figure, 12 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Role-Play Prompting in LLM
  4. Methods
  5. Problem Definition
  6. Persona-Agnostic Semantic Invariance.
  7. Persona Switch
  8. Output Candidate Generation.
  9. Output Confidence Comparison.
  10. Answer Extraction.
  11. Experiments
  12. Experimental setup
  13. Experimental Settings.
  14. Results and analysis
  15. Main Results.
  16. ...and 24 more sections

Figures (1)

  • Figure 1: Relative Correct Rates Show No Clear Superiority Between Prompting Strategies. Each bar represents the proportion of instances where only one strategy answered correctly, highlighting that neither one demonstrates clear superiority.