Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning

TL;DR

This paper introduces Role-Aware Reasoning (RAR), a method to improve role-playing agents by maintaining character consistency and reasoning style through explicit role guidance and style optimization, enhancing their internal thought processes.

Contribution

The paper proposes a novel RAR approach with RIA and RSO stages to address attention diversion and style drift in RPAs, improving their internal reasoning and character fidelity.

Findings

RAR significantly improves RPA performance

Effective mitigation of attention diversion

Enhanced character style consistency

Abstract

The advancement of Large Language Models (LLMs) has spurred significant interest in Role-Playing Agents (RPAs) for applications such as emotional companionship and virtual interaction. However, recent RPAs are often built on explicit dialogue data, lacking deep, human-like internal thought processes, resulting in superficial knowledge and style expression. While Large Reasoning Models (LRMs) can be employed to simulate character thought, their direct application is hindered by attention diversion (i.e., RPAs forget their role) and style drift (i.e., overly formal and rigid reasoning rather than character-consistent reasoning). To address these challenges, this paper introduces a novel Role-Aware Reasoning (RAR) method, which consists of two important stages: Role Identity Activation (RIA) and Reasoning Style Optimization (RSO). RIA explicitly guides the model with character profiles during reasoning to counteract attention diversion, and then RSO aligns reasoning style with the character and scene via LRM distillation to mitigate style drift. Extensive experiments demonstrate that the proposed RAR significantly enhances the performance of RPAs by effectively addressing attention diversion and style drift.