AdaMARP: Adaptive Multi-Agent Framework for Immersive Role-Playing
Researchers have introduced AdaMARP, a novel framework designed for adaptive multi-agent interactions in immersive role-playing scenarios, which aims to overcome the shortcomings of current LLM-based systems, particularly in terms of immersion and adaptability. This framework features an innovative message format that combines [Thought], (Action), <Environment>, and Speech, alongside a dedicated Scene Manager that directs role-playing through specific actions (init_scene, pick_speaker, switch_scene, add_role, end), each with justifications. To develop these functionalities, the team created AdaRPSet for the Actor Model and AdaSMSet for managing orchestration choices, as well as AdaptiveBench for evaluating trajectories. Experiments conducted with various backbones validate the framework's efficacy.
Key facts
- AdaMARP is an adaptive multi-agent role-playing framework.
- It features an immersive message format with [Thought], (Action), <Environment>, and Speech.
- A Scene Manager governs role-playing via discrete actions: init_scene, pick_speaker, switch_scene, add_role, end.
- AdaRPSet and AdaSMSet are constructed for training the Actor Model and orchestration decisions.
- AdaptiveBench is introduced for trajectory-level evaluation.
- Experiments are conducted across multiple backbones.
- The framework aims to improve immersion and adaptability in LLM role-playing.
- Existing systems under-model dynamic environmental information and assume static scenes.
Entities
Institutions
- arXiv