ARdena Enables Scenario-Driven Control for Real-Time LLM Agents.

While large language models (LLMs) have significantly advanced conversational agent capabilities, maintaining reliable control over their behavior in real-time interactive environments remains a considerable hurdle. Existing methods often rely on fine-tuning or alignment procedures, which are difficult to adapt quickly to evolving interaction requirements. This paper presents a solution through layered scenario-driven LLM control. This innovative framework enables runtime behavior modification using structured prompting. By integrating persistent contextual information with specific constraints tailored to different scenarios, the agent's actions can be altered during an interaction without needing to change the underlying LLM itself. This offers a flexible and agile approach to managing agent behavior. The framework is embodied in ARdena, a real-time multimodal agent that combines speech interaction, visual perception, tool utilization, and avatar-based response generation. Evaluations of ARdena confirm the effectiveness of scenario definitions in producing diverse interaction behaviors while maintaining stable, real-time operation. This highlights the power of scenario-driven prompting for precise control over LLM agents.