AI role-playing is evolving from simple "text-based conversations" to complex "immersive dramas." Recently, Zhejiang University in collaboration with Tencent Youtu Lab proposed a self-adaptive multi-agent interaction framework called AdaMARP. This research aims to address the current pain points of large language models in role-playing, such as the lack of environmental awareness and rigid storytelling, endowing AI with the scene management and narrative abilities akin to a "director." Currently, this achievement has been accepted by the international academic conference ACL 2026.

Core Pain Points: Missing "Environment" and "Director"

In existing AI role-playing scenarios, users can converse with historical or literary characters, but interactions are often limited to text-based exchanges, with static settings and characters. For example, in a detective scenario, traditional AI systems often act like "repeating talking machines," unable to reason based on environmental clues (such as wax stains on the carpet) or handle complex narrative demands like multiple character turns or scene transitions. This "empty" mode makes it difficult for users to experience real authenticity and narrative tension.

AdaMARP Framework: Four-Channel Messages and Dynamic Scheduling

To break the deadlock, the research team designed a new interaction logic. First, AdaMARP introduces a "four-channel message format," breaking down each round of interaction into "Thought - Action - Environment - Speech." In this model, AI no longer just outputs dialogue but also interweaves the atmosphere of the environment (such as flickering gas lamps), internal considerations, and body language, forming a complete chain of causality.

Second, the framework introduces a "Scene Manager" role, acting as the "director" of the entire narrative. It possesses five core capabilities: initializing the scene, selecting the speaker, switching scenes, dynamically introducing new characters, and ending the interaction. This means the AI system can autonomously decide when to move from the crime scene to a witness's home, or when to have a new suspect "walk through the door."

Training and Evaluation: From Literature to Simulation

To enable AI to truly master "acting" and "directing" abilities, the research team built high-quality datasets AdaRPSet and AdaSMSet. These datasets not only include deep character profiles and interaction trajectories extracted from 81 classic literary works but also cover 20 different thematic synthetic plots, ensuring that the model learns the texture of literature while mastering the logic of dynamic scheduling.

In addition, the team introduced a complementary evaluation framework called AdaptiveBench. Unlike traditional single-turn conversation evaluations, this framework scores models at the trajectory level, focusing on character consistency, environmental perception, and the naturalness of narrative progression, thereby comprehensively assessing AI performance in complex long-text interactions.

This framework provides a new technical path for immersive interactive scenarios such as detective reasoning and adventure storytelling. Through the deep coupling of environment and narrative logic, AI is evolving from a mere chat assistant into a digital performer with advanced creative consciousness.