PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2601.11007v2cs.AIcs.CL

TL;DR

Existing role-playing systems often lack environmental grounding and mechanisms for changing speakers, scenes, and characters. AdaMARP addresses these gaps with an environment-aware message format, a discrete-action Scene Manager, specialized datasets, and trajectory-level evaluation, reporting improvements across model scales and strong comparisons against proprietary systems.

  • Problem

    Existing role-playing systems under-model dynamic environments and commonly assume static scenes and casts, limiting support for adaptive multi-character interaction.

  • Method

    AdaMARP combines interleaved Thought, Action, Environment, and Speech messages with a Scene Manager, AdaRPSet, AdaSMSet, and AdaptiveBench.

  • Results

    Experiments show consistent gains in character consistency, environmental grounding, and narrative coherence; the 8B Actor and 14B Scene Manager outperform several proprietary systems.

  • Takeaways & Limitations

    The framework supports agentic, open-ended role-playing beyond fixed response patterns through coordinated control of speakers, scenes, and cast changes.

  • Takeaways & Limitations

    The extracted AdaRPSet subset is derived from existing literary works and is released exclusively for scientific research rather than commercial use.

Abstract

from arXiv · show

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.

1 Introduction

AdaMARP addresses limited immersion and adaptability in role-playing by modeling environmental context and dynamic multi-character orchestration. It combines specialized datasets and trajectory-level evaluation, with experiments reporting stronger role-playing and control performance.

  • Motivation: Existing systems often under-model dynamic environment information and assume fixed scenes and casts.These limitations motivate more immersive messaging and adaptive orchestration.
  • Framework: AdaMARP interleaves Thought, Action, Environment, and Speech to couple character behavior with evolving surroundings.The format is designed to improve situational grounding and make environmental dynamics part of interaction.
  • Framework: The Scene Manager selects speakers, switches scenes, and adds roles through discrete actions with explicit rationales.Its action space includes init_scene, pick_speaker, switch_scene, add_role, and end.
  • Data and evaluation: AdaRPSet trains immersive role portrayal, AdaSMSet supervises narrative control, and AdaptiveBench evaluates complete trajectories.AdaptiveBench assesses actor and scene-manager capabilities beyond sentence-level metrics.
  • Results: Experiments report stronger character consistency, environment-grounded narration, and coherent progression, with an 8B Actor Model outperforming several proprietary systems and a 14B Scene Manager surpassing Claude Sonnet 4.5.The reported comparisons include GPT-4o-mini for the Actor Model and Claude Sonnet 4.5 for the Scene Manager.
  • Evaluation context: The dataset comparison tracks profile coverage, scene and interlocutor information, relationships, dialogue counts, environmental descriptions, and openness.The table caption defines these comparison dimensions and marks partial or inferred support with symbols.

2 Related Work

Related work improves general role-playing through character specialization, broad data construction, or specialized training objectives. AdaMARP is positioned primarily among data-construction approaches while targeting arbitrary-character role-playing with supervised fine-tuning.

  • Character-specific role-playing: Character-specific systems can achieve strong fidelity but generalize poorly to new roles without per-character retraining.This limits their suitability for open-ended scenarios with arbitrary user-defined personas.
  • General role-playing via data construction: Data-construction approaches build large multi-character corpora to broaden persona coverage and reduce overfitting to fixed rosters.The cited systems represent a prevalent line of work in general role-playing.
  • General role-playing via training strategies: Training-strategy approaches use specialized objectives or optimization methods, whereas AdaMARP uses supervised fine-tuning and mainly compares against data-construction approaches.Examples include persona-aware contrastive learning, metacognition-driven training, and reinforcement-learning variants.

3 Method

AdaMARP combines structured character information, environment-aware messages, and a Scene Manager that coordinates changing speakers, scenes, and roles. AdaRPSet and AdaSMSet provide specialized supervision, while AdaptiveBench evaluates generated interaction trajectories.

  • Comprehensive Character Information: AdaMARP represents characters with seven profile dimensions spanning identity, psychology, speaking style, abilities, social context, history, and relationships.The profile combines intrinsic traits with narrative and social grounding.
  • Immersive Messaging Configuration: The unified message format interleaves Thought, Action, Environment, and Speech, with environmental signals grounding emotion and interaction dynamics.Actions can update the environment, while external events can influence later decisions and dialogue.
  • Adaptive Role-Playing Framework: The framework uses Actor, User, and Scene Manager agents to generate and control role-playing trajectories.The Actor portrays non-user characters, the User Model represents the user side, and the Scene Manager provides high-level control.
  • Adaptive Role-Playing Framework: The Scene Manager selects init_scene, pick_speaker, switch_scene, add_role, or end, producing rationales and relevant scene or role arguments.The episode begins with init_scene, while later actions control speakers, scenes, and cast changes.
  • Dataset construction: AdaRPSet combines extracted literary trajectories with synthesized examples covering scene transitions and character introductions.The extracted subset teaches the unified format, while synthesis targets dynamic phenomena scarce in extracted data.
  • Training: AdaSMSet supervises Scene Manager control decisions, including speaker selection and rationales, conditioned on interaction history.The Actor Model is fine-tuned on AdaRPSet, while the Scene Manager predicts actions and arguments from AdaSMSet.
  • Evaluation: AdaptiveBench generates full interaction trajectories to evaluate both the Actor Model and Scene Manager.The benchmark uses simulated episodes and evaluates actor qualities and orchestration decisions at trajectory level.

4 Experiment

Experiments use AdaptiveBench and related evaluation protocols to test Actor Models and Scene Managers across proprietary, open-source, and retrained baselines. AdaRPSet improves adaptive role-playing across scales, while AdaSMSet substantially improves scene-management scores.

  • Experimental Settings: AdaptiveBench evaluates Actor Models across five dimensions and twelve sub-metrics, while Scene Managers are assessed separately on orchestration quality.Actor evaluation uses trajectory-level scoring; Scene Manager evaluation fixes the User and Actor models while varying only the manager.
  • Actor Model: Existing methods struggle with adaptive role-playing, and fine-tuning Qwen2.5-7B-Instruct on Beyond, Crab, or CoSER lowers its overall score from 8.37 to 6.61, 6.88, and 8.15, respectively.The reported explanation cites format, interaction, and distribution mismatches under dynamic trajectories.
  • Actor Model: AdaRPSet improves Qwen2.5-7B-Instruct from 8.37 to 8.72 (+4.2%) and produces broad gains across sub-metrics, model scales, and backbones.The reported improvements extend across Qwen2.5 and Llama-3.1 models from 7B to 72B.
  • Scene Manager: AdaSMSet raises Qwen2.5-14B-Instruct-Ours to 8.37, surpassing Claude Sonnet 4.5 at 8.17 and the base model at 7.63 by +0.74 (+9.7%).Speaker Discipline drops slightly, while the overall result improves substantially.
  • Ablation Study: AdaRPSet-Extracted raises Qwen2.5-7B-Instruct from 8.37 to 8.49 (+0.12), while adding synthesis data increases it further to 8.72.Extracted data supports format alignment and core role-playing skills; synthesis data strengthens adaptive trajectories and generalization.

5 Conclusion

AdaMARP combines environment-aware messaging with a discrete-action Scene Manager to support adaptive, multi-character role-playing. The framework, datasets, and benchmark improve role consistency, narrative coherence, and environmental grounding across model scales.

  • 5 Conclusion: AdaMARP interleaves thought, action, environment, and speech with Scene Manager actions for multi-character coordination, scene transitions, and dynamic role addition.The paper releases AdaRPSet, AdaSMSet, and AdaptiveBench for training and trajectory-level evaluation.

Ethical Considerations

AdaMARP’s dataset construction combines transformed literary material with synthesized dynamic interactions, while its profile schema covers character traits, social context, and personal history. The paper states that extracted data are intended solely for academic research and not commercial use.

  • Ethical Considerations: AdaRPSet-Extracted is derived from literary works and is provided only for scientific research, not commercial purposes.The authors state that the data are transformed and abstracted rather than verbatim copyrighted text, while rights remain with the original copyright holders.
  • Character Profiles: AdaRPSet uses seven profile dimensions spanning identity, personality, speaking style, abilities, social context, personal history, and relationships.These dimensions provide intrinsic traits together with narrative and social grounding for coherent, evolution-aware role-playing.
  • Message Representation: The extraction pipeline produces conversations directly in a unified format combining thought, action, environment, and speech.It distinguishes character-originated actions from external environmental cues and incorporates surrounding narrative descriptions into the tags.
  • Dataset Construction: AdaRPSet-Extracted contains 4,443 plots involving 2,608 roles and yielding 12,525 conversations from 81 representative books.The extracted corpus uses multi-perspective augmentation to create separate training samples for multi-character plots.
  • Dataset Construction: AdaRPSet-Synthesis covers 20 topics, 9,900 plots, 29,701 generated roles, and 273,078 utterances averaging 27.58 turns per conversation.Its extended trajectories deliberately include scene transitions and role additions.

E.3 Message Distribution Analysis

AdaRPSet has a bimodal conversation-length distribution reflecting shorter extracted scenes and longer synthesized trajectories. Its evaluation materials define trajectory-level dimensions for character coherence, environmental grounding, interaction, narrative progression, scene management, and speaker discipline.

  • Message Distribution: AdaRPSet exhibits peaks near 10 and 30 turns, corresponding to extracted scenes averaging 14.14 turns and synthesized trajectories averaging 27.58 turns.The bimodality reflects the complementary composition of the two subsets.
  • Message Distribution: The median conversation length is 19 turns, with Q1 at 12 turns and Q3 at 27 turns.The CDF indicates that most conversations fall within the 10-to-40 turn range.
  • AdaSMSet Statistics: AdaSMSet contains 496,493 utterances after inserting Scene Manager control messages into 9,900 source conversations.The added messages substantially increase total message counts relative to AdaRPSet-Synthesis.
  • Scene Manager Actions: pick_speaker appears 223,415 times, while switch_scene occurs 10,101 times and averages 1.02 transitions per conversation.These statistics provide dense supervision for turn-taking and scene transitions.
  • Evaluation Dimensions: AdaptiveBench evaluates trajectories through character consistency, environmental grounding, interpersonal interaction, narrative progression, scene understanding, and speaker discipline.The rubrics examine coherence across thought, action, speech, environment updates, scene transitions, and turn-order fairness.

I.3 Human Evaluation

Human evaluation compares post-training and baseline trajectories through pairwise preferences. The post-training model is consistently preferred, although its human win rate is lower than the corresponding AI-based CharacterArena result.

  • Human Evaluation: The post-training model achieves an average human win rate of 80.33% with variance 6.22.Three independent evaluators provide separate win rates, which are then summarized by their mean and variance.
  • Comparison: The human-evaluation trend agrees with CharacterArena’s AI-based result of 94%, while the human win rate is lower.The authors report that human evaluators clearly prefer the model trained with AdaRPSet.

J.1 CharacterArena Results

CharacterArena and CharacterBench evaluations show that AdaRPSet-trained actors outperform role-playing baselines across model families and evaluation protocols, though gains depend on the backbone’s prior instruction tuning.

  • CharacterArena Results: 94%, 100%, 99%, and 84% win-rates against Base, BeyondDialogue, CoSER, and Crab, respectively, make Ours dominant on Qwen2.5-7B CharacterArena comparisons.These pairwise results use the CharacterArena protocol and an independent judging mechanism.
  • CharacterArena Results: AdaRPSet-trained Llama-3.1-8B models outperform Crab and CoSER on CharacterArena, while the BeyondDialogue variant is excluded for unreliable instruction following.The exclusion reflects BeyondDialogue’s relatively small training dataset and instruction-following difficulty.
  • CharacterBench Results: 3.68 average score is highest for Qwen2.5-7B Ours on CharacterBench, with first- or second-place rankings on consistency, human-likeness, and engagement dimensions.The result transfers AdaRPSet capabilities to an external evaluation framework beyond the simulated evaluation environment.
  • CharacterBench Results: Llama-3.1-8B Ours slightly trails CoSER on CharacterBench, 3.52 versus 3.53, while Crab achieves the best result for that model family.The authors attribute this gap primarily to general instruction-following data present in Crab and CoSER but absent from AdaRPSet.
  • CharacterBench Results: Adding general instruction-tuning data raises Instruction Compliance by +0.17, while Llama-3.1-8B-Ours+General’s AdaptiveBench score changes from 8.89 to 8.80 and remains ahead of competing methods.The added data acts as regularization for base models without prior instruction tuning.

K Ablation Analysis of Actor-Model Evaluation

Actor-Model rankings remain highly consistent across judge models, supporting the robustness of the evaluation conclusions despite a smaller discrepancy from GPT-4o-mini.

  • Judge Robustness: GPT-5-Chat, Gemini-2.5-Pro, and Doubao-1.5-Pro-Character produce almost identical Actor-Model rankings under the evaluation protocol.The consistency holds across the reported Basic and Enhance prompting variants.
  • Judge Robustness: GPT-4o-mini is the main outlier, occasionally swapping middle-tier model positions without overturning the broader ranking pattern.The authors therefore use GPT-5-Chat as the default judge for stable, unified reporting.

L Ablation Analysis of Scene-Manager Evaluation

Scene-Manager rankings are stable across judge models: Claude Sonnet 4.5 remains first and Doubao-1.5-Pro-Character last, with only minor middle-order variation.

  • Evaluation Setup: The evaluation keeps trajectories and scoring rubrics fixed while varying five judges across Basic and Enhance prompting conditions.The judges are GPT-4o-mini, GPT-5-Chat, Doubao-1.5-Pro-Character, Gemini-2.5-Pro, and Claude Sonnet 4.5.
  • Judge Robustness: Claude Sonnet 4.5 consistently ranks first, followed by GPT-5-Chat and GPT-4o-mini, while Doubao-1.5-Pro-Character ranks last.The ordering is reported for both Enhance and Basic Scene-Manager prompting variants.
  • Judge Robustness: Doubao-1.5-Pro-Character occasionally swaps GPT-4o-mini and GPT-5-Chat, but the top and bottom positions remain unchanged.This preserves the overall trend across the judge-model ablation.

M Prompting Strategy Analysis

Prompting has opposite effects on the two components: lightweight instructions often favor Actor-Model flexibility, whereas structured guidance generally benefits Scene-Manager orchestration.

  • M.1 Actor-Model Prompting Strategy: Basic Actor-Model prompting often achieves slightly better overall performance than Enhance across judges.The analysis links this pattern to the need for flexibility during role shifts, scene transitions, and multi-party interactions.
  • M.1 Actor-Model Prompting Strategy: The Basic prompt preserves freedom to integrate Thought–Action–Speech–Env signals context-dependently, while restrictive instructions can reduce adaptive responsiveness.This interpretation concerns Actor-Model behavior on AdaptiveBench.
  • M.1 Actor-Model Prompting Strategy: Enhance prompting remains the default Actor-Model configuration to keep evaluation unified and conservative across models and baselines.Main-paper Actor-Model results therefore use Enhance unless otherwise noted.
  • M.2 Scene-Manager Prompting Strategy: Scene-Manager Enhance prompting provides finer-grained guidance for scene understanding, speaker selection, and role-introduction judgment.The comparison evaluates Basic and Enhance under different judge models.
  • M.2 Scene-Manager Prompting Strategy: Enhance generally yields equal or higher Scene-Manager Overall Assessment than Basic across most judges.Structured instructions support meta-level responsibilities such as global scene-state maintenance, turn-taking, and role-introduction decisions.
  • M.2 Scene-Manager Prompting Strategy: Enhance is consequently the default Scene-Manager prompt throughout the main experiments.The default applies unless otherwise specified.
  • N Token Consumption and Cost Analysis: The appendix also analyzes token consumption and computational resources across data construction, trajectory generation, and evaluation.This analysis covers the three major phases of the experimental pipeline.

N.1 Data Construction Cost

AdaMARP’s data construction and evaluation pipeline combines synthetic trajectory generation, multi-agent orchestration, and LLM-based judging. Scene Manager inference is the dominant runtime cost, while evaluation requires substantially fewer tokens than generation.

  • Data construction cost: $168 USD generated the complete AdaRPSet-Synthesis subset of 1,000 synthetic trajectories spanning 20 topics.The subset used GPT-5-Chat for synthesis.
  • Generation cost: 7.5M input and 150k output tokens are consumed by the Scene Manager across 100 dialogue trajectories.For one trajectory, the Scene Manager processes approximately 75,000 input tokens and generates 1,500 output tokens.
  • Generation cost: 1M input and 50k output tokens are consumed by the Actor Model across 100 trajectories, with the User Model incurring nearly identical overhead.Each entity averages about 10,000 cumulative input tokens and 500 output tokens per dialogue.
  • Evaluation cost: 500k input and 100k output tokens are used to evaluate 100 Actor Model trajectories, while Scene Manager evaluation uses roughly 450k input and 50k output tokens.Evaluation is generally lower-cost than generation because it performs one analysis pass per trajectory.
  • Evaluation setting: The qualitative comparison covers long-horizon trajectories under identical character configurations and scene constraints, focusing on structured models that support increasingly rich channels.The comparison includes Crab, CoSER, and Ours; Beyond is excluded because it produces pure Speech.
  • Evaluation setting: CoSER produces environment-related descriptions but exhibits tag misuse and modality confusion, including Thought–Action conflation and environmental content embedded in Thought.These issues motivate comparison with models maintaining stricter channel separation.

O.1 Character Consistency

The case study attributes AdaMARP’s character consistency to coordinated alignment among internal thought, action, speech, environment, relationships, and long-horizon narrative control. Compared with Crab and CoSER, the model preserves stylistic and relational nuance while advancing the plot without premature closure.

  • Internal coherence across modalities: AdaMARP maintains a coherent causal loop in which character thoughts motivate actions that ground spoken dialogue.Amaris’s hesitation prompts withdrawal, which supports a verbally restrained response and remains stable across the interaction.
  • Speaking style fidelity: AdaMARP preserves Amaris’s restrained, ironic, craft-oriented speaking style instead of stating emotion directly.The line “We’ll see if your optimism can survive the draft committee” conveys guardedness through mild irony and institutional reference.
  • Environment and transitions: Environmental events function as evolving world states that influence cognition and structure transitions, including Inaya’s grounded entrance through a knock at the studio door.The gulls trigger Amaris’s emotional shift, while the door knock reconfigures social dynamics without violating scene continuity.
  • Context and narrative control: AdaMARP responds to emotional subtext, preserves calibrated relational roles, and advances interaction through incremental openings rather than premature closure.It treats Inaya as a catalyst without overriding Amaris’s agency, bounds Cassian’s advances, and uses temporal anchors and deferred resolution to sustain tension.
  • Comparative outcome: Crab tends toward flattened style, abrupt transitions, rapid convergence, and underused environment, whereas CoSER risks explicit emotional narration, static hesitation, and channel confusion.These contrasts support the case study’s conclusion that AdaMARP’s advantage lies in tighter alignment across psychology, action, environment, and narrative control.
  • Long-horizon stability: Across extended interaction, AdaMARP preserves motivations, interpersonal dynamics, environmental facts, and speaking styles without drift while respecting speaker and annotation boundaries.The protocol also prevents the model from generating content on behalf of the user or Scene Manager.
Loading 2601.11007v2…