Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AdaMARP: An Adaptive Multi-Agent Interaction Framework for General Immersive Role-Playing

Zhenhua Xu, Dongsheng Chen, Shuo Wang, Jian Li, Chengjie Wang, Meng Han, Yabiao Wang

TL;DR

AdaMARP presents a general adaptive framework for immersive role-playing by coupling an environment-aware messaging format with a discrete Scene Manager that orchestrates multi-character interaction and scene transitions. It introduces AdaRPSet and AdaSMSet to supervise Actor and Scene Manager behaviors, respectively, and AdaptiveBench to evaluate full trajectories rather than single-turn outputs. Across multiple backbones and scales, AdaRPSet improves character consistency and environmental grounding, with an 8B Actor model outperforming several commercial LLMs, while AdaSMSet enables smoother scene transitions and more natural role introductions, surpassing Claude Sonnet 4.5 at 14B. The framework’s trajectory-level evaluation and data releases support robust, generalizable immersive role-playing beyond static prompts or fixed rosters, advancing practical open-ended agentic narratives.

Abstract

LLM role-playing aims to portray arbitrary characters in interactive narratives, yet existing systems often suffer from limited immersion and adaptability. They typically under-model dynamic environmental information and assume largely static scenes and casts, offering insufficient support for multi-character orchestration, scene transitions, and on-the-fly character introduction. We propose an adaptive multi-agent role-playing framework, AdaMARP, featuring an immersive message format that interleaves [Thought], (Action), <Environment>, and Speech, together with an explicit Scene Manager that governs role-playing through discrete actions (init_scene, pick_speaker, switch_scene, add_role, end) accompanied by rationales. To train these capabilities, we construct AdaRPSet for the Actor Model and AdaSMSet for supervising orchestration decisions, and introduce AdaptiveBench for trajectory-level evaluation. Experiments across multiple backbones and model scales demonstrate consistent improvements: AdaRPSet enhances character consistency, environment grounding, and narrative coherence, with an 8B actor outperforming several commercial LLMs, while AdaSMSet enables smoother scene transitions and more natural role introductions, surpassing Claude Sonnet 4.5 using only a 14B LLM.

AdaMARP: An Adaptive Multi-Agent Interaction Framework for General Immersive Role-Playing

TL;DR

AdaMARP presents a general adaptive framework for immersive role-playing by coupling an environment-aware messaging format with a discrete Scene Manager that orchestrates multi-character interaction and scene transitions. It introduces AdaRPSet and AdaSMSet to supervise Actor and Scene Manager behaviors, respectively, and AdaptiveBench to evaluate full trajectories rather than single-turn outputs. Across multiple backbones and scales, AdaRPSet improves character consistency and environmental grounding, with an 8B Actor model outperforming several commercial LLMs, while AdaSMSet enables smoother scene transitions and more natural role introductions, surpassing Claude Sonnet 4.5 at 14B. The framework’s trajectory-level evaluation and data releases support robust, generalizable immersive role-playing beyond static prompts or fixed rosters, advancing practical open-ended agentic narratives.

Abstract

LLM role-playing aims to portray arbitrary characters in interactive narratives, yet existing systems often suffer from limited immersion and adaptability. They typically under-model dynamic environmental information and assume largely static scenes and casts, offering insufficient support for multi-character orchestration, scene transitions, and on-the-fly character introduction. We propose an adaptive multi-agent role-playing framework, AdaMARP, featuring an immersive message format that interleaves [Thought], (Action), <Environment>, and Speech, together with an explicit Scene Manager that governs role-playing through discrete actions (init_scene, pick_speaker, switch_scene, add_role, end) accompanied by rationales. To train these capabilities, we construct AdaRPSet for the Actor Model and AdaSMSet for supervising orchestration decisions, and introduce AdaptiveBench for trajectory-level evaluation. Experiments across multiple backbones and model scales demonstrate consistent improvements: AdaRPSet enhances character consistency, environment grounding, and narrative coherence, with an 8B actor outperforming several commercial LLMs, while AdaSMSet enables smoother scene transitions and more natural role introductions, surpassing Claude Sonnet 4.5 using only a 14B LLM.
Paper Structure (122 sections, 1 equation, 3 figures, 39 tables, 2 algorithms)

This paper contains 122 sections, 1 equation, 3 figures, 39 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Character-Specific Role-Playing.
  4. General Role-Playing via Data Construction.
  5. General Role-Playing via Training Strategies.
  6. Method
  7. Design Principles of AdaMARP
  8. Comprehensive Character Information
  9. Immersive Messaging Configuration
  10. Adaptive Role-Playing Framework
  11. Dataset Curation
  12. Dataset for Actor Model
  13. Dataset for Scene Manager
  14. Actor Model and Scene Manager Training
  15. AdaptiveBench: Evaluation Framework
  16. ...and 107 more sections

Figures (3)

  • Figure 1: Overall framework of AdaMARP. The left part illustrates the construction of AdaRPSet and AdaSMSet and the components of a comprehensive role profile. The upper middle part shows the roles played by the three agents, while the lower middle part depicts the dialogue trajectory generation under the orchestration of the scene manager, given predefined roles and an initial scenario. The right part presents the evaluation of the actor models and the scene manager based on the generated trajectories.
  • Figure 2: Word cloud visualizations of six key profile dimensions in AdaRPSet. The layout displays: (a) Identity & Appearance, (b) Personality & Psychology, (c) Speaking Style, (d) Abilities, Interests & Achievements, (e) Social & Historical Context, and (f) Personal History Arc. The diversity of terms indicates a broad coverage of character archetypes and backgrounds.
  • Figure 3: Distribution analysis of message counts per conversation in AdaRPSet. The figure aggregates (a) Histogram, (b) KDE, (c) Boxplot, and (d) CDF. The bimodal nature of the distribution reflects the complementary characteristics of the Extracted and Synthesis subsets.