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

TL;DR

This work tackles the problem of RPAs lacking deep, human-like internal thinking by introducing Role-Aware Reasoning (RAR), which combines Role Identity Activation (RIA) to anchor reasoning in a character’s core traits and Reasoning Style Optimization (RSO) to adapt the thinking style to scene context. By distilling reasoning traces from a large reasoning model into a smaller LLM, RAR mitigates attention diversion and style drift, enabling more consistent and believable character portrayals. Extensive experiments on RoleBench-derived data and standard RP A benchmarks (CharacterBench and SocialBench) show that RAR improves persona fidelity, knowledge recall, and social judgment, with ablations confirming the complementary roles of RIA and RSO. The results suggest that explicit guidance of internal reasoning is a promising direction for enhancing complex generative tasks like role-playing, with potential extensions to finer character attributes, long-term memory, and larger teacher models.

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.

Figures (10)

• Figure 1: Overview of the proposed RAR. Given a user query, a LRM generates structured thoughts. However, traditional reasoning may suffer from attention diversion and style drift, leading to generic, out-of-character responses. To address this, our method incorporates RIA and RSO. RIA activates key role traits (e.g., emotions, motivations) to distill role-consistent thoughts. RSO guides the model to generate reasoning traces in suitable styles depending on context, enabling dynamic control over logical and narrative expression.
• Figure 2: Analysis of RIA components' impact on CharacterBench persona metrics. The figure shows the performance of the full RAR model versus variants where specific elements of the Role Identity Activation (Emotion, Experience, Standpoint, Motivation) are individually ablated.
• Figure 3: The t-SNE visualization of hidden states from different layers of the RAR model for responses generated under different reasoning style prompts (e.g., fact-based vs. character-knowledge-based).
• Figure 4: The prompt for RIA.
• Figure 5: The prompt for logical scenarios of the RSO.