Dynamic Context Adaptation for Consistent Role-Playing Agents with Retrieval-Augmented Generations - Abstract

Sep 18, 2026

Uploaded by Logic

Building role-playing agents (RPAs) that faithfully emulate specific characters remains challenging because collecting character-specific utterances and continually updating model parameters are resource-intensive, making retrievalaugmented generation (RAG) a practical necessity. However, despite the importance of RAG, there has been little research on RAG-based RPAs. When a persona lacks knowledge relevant to a query, RAG-based RPAs hallucinate, undermining response accuracy. The paper proposes AMADEUS, a training-free framework improving persona consistency even for questions beyond character knowledge, and introduces CHARACTERRAG with 15 fictional characters and Q&A pairs, with code and data to be released.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages14

Reading Time35 min

AccessFree to download

Abstract

1 Introduction

Retrieval-augmented role-playing agents and challenges

Proposed AMADEUS framework

CHARACTERRAG dataset

Evaluation and release

Dynamic Context Adaptation for Consistent Role-Playing Agents with Retrieval-Augmented Generations Jeiyoon Park1 , Yongshin Han1 , Minseop Kim1 , Kisu Yang2 , 1 SOOP, 2 Korea University {naruto, winterfeb, [usio}@sooplive.com](usio}@sooplive.com), [willow4@korea.ac.kr](willow4@korea.ac.kr) [https://huggingface.co/datasets/naruto-soop/CharacterRAG](https://huggingface.co/datasets/naruto-soop/CharacterRAG) arXiv :2508 .020 16v4 [ cs .AI ] 5 Feb 2026 Abstract Building role-playing agents (RPAs) that faithfully emulate specific characters remains challenging because collecting character-specific utterances and continually updating model parameters are resource-intensive, making retrievalaugmented generation (RAG) a practical necessity. However, despite the importance of RAG, there has been little research on RAG-based RPAs. For example, we empirically find that when a persona lacks knowledge relevant to a given query, RAG-based RPAs are prone to hallucination, making it challenging to generate accurate responses. In this paper, we propose AMADEUS, a training-free framework that can significantly enhance persona consistency even when responding to questions that lie beyond a character’s knowledge. In addition, to underpin the development and rigorous evaluation of RAG-based RPAs, we manually construct CHARACTERRAG, a role-playing dataset that consists of persona documents for 15 distinct fictional characters totaling 976K written characters, and 450 question–answer pairs. We find that our proposed method effectively models not only the knowledge possessed by characters, but also various attributes such as personality. The code and dataset will be made publicly available at our GitHub. 1 Introduction Large language models with long-context capabilities are engineered to manage lengthy input sequences, allowing them to interpret and utilize extended contextual information (OpenAI, 2025 ; Qwen et al., 2025 ; Gemini Team, 2025, 2024) . Although LLMs exhibit enhanced abilities in understanding extended contexts, they still face significant challenges when handling tasks involving genuinely long contexts (Li et al., 2024a) . Furthermore, utilizing all relevant information from longcontext models to answer each query can be computationally expensive (Li et al., 2024b) . Retrieval- |
Roleplay Prompt
From now on, you are Tanjiro Kamado. Refer to the given context and Tanjiro Kamado’s attributes, and answer the user’s question as Tanjiro Kamado. | | | | --- | --- | --- | |
A question beyond the Character's knowledge | | | | “Do you feel comfortable approaching someone you’re interested in and starting a conversation?”
| | | |
Context (Naïve RAG) | |
Dynamic Context (Ours) | | | | | | -When his favorite food, angelica tree shoots \# Other: (tara-no-me) , is placed in front of him, he -Even before his relationships get close, he pretends to be indifferent and tries to keep a calm has a habit of calling people by their names demeanor (It seems that because he sees himself very quickly. In Tanjiro's case, he calls others by as the eldest son, he doesn't want to show how name much faster than is typical for Japanese excited he is about food. ) . However, his nostrils people. ( ) flare, so everyone around him can already tell how he really feels. ( )

Tanjiro’s Response (Hallucinated)

“Approaching someone I'm interested in and starting a conversation just comes naturally to me. Of course, if I'm trying to hide my feelings, like when I'm in front of a durup (angelica-tree shoots), it can get a little awkward... I guess my feelings always show—like when my nostrils flare up! Haha.”

Tanjiro’s Response (Ours)

“Yeah, that’s right. For me, it actually feels more natural to approach someone first when I meet them, call their name, and start a conversation. I think it’s easy to become friends, even with someone I’m meeting for the first time

FAQs

Why are retrieval-augmented generation (RAG) methods needed for role-playing agents?

Collecting character-specific utterances and continually updating model parameters is resource-intensive, so RAG provides an efficient way to ground responses with external information.

What problem is observed with naive RAG-based RPAs?

When the persona lacks knowledge relevant to a query, naive RAG tends to use less relevant chunks, leading to hallucinated and inaccurate responses.

What does AMADEUS contribute to persona consistency?

AMADEUS is a training-free framework that significantly enhances persona consistency even when answering questions that lie beyond a character’s knowledge.

Sep 18, 2026

Uploaded by Logic