Cinema Multiverse Lounge: Enhancing Film Appreciation via Multi-Agent Conversations

Research Background and Problems

• What issues or challenges did the authors identify?
  Traditional methods of film appreciation and discussion primarily rely on face-to-face communication or digital platforms, which are limited by physical distance, the availability of discussion partners, and the depth of topics. Additionally, existing interactive AI platforms (e.g., Character.ai) typically focus on interactions with a single agent, lacking comprehensive discussions across multiple perspectives.

• Why is this issue important?
  Film appreciation is not only a form of entertainment but also an experience involving deep reflection and emotional resonance. Film discussions and interpretations can help viewers deepen their understanding of narratives and artistic expression. However, traditional and current digital methods fail to fully support a rich, multi-perspective film appreciation experience.

• Research Motivation and Related Work
  With the rapid development of large language models (LLMs), researchers have the opportunity to design systems that support multi-agent interactions, thereby enhancing viewers' understanding and enjoyment of films through multi-perspective dialogues. Existing research primarily focuses on single-character simulations or task-oriented multi-agent systems, but lacks applications in the domain of film appreciation.

Solution

• What methods or solutions did the authors propose?
  The authors proposed a multi-agent conversational system named Cinema Multiverse Lounge (CML), enabling users to interact in real-time with various film-related characters driven by LLMs, including movie characters, filmmakers, and virtual audiences.

• What are the innovative aspects of this solution?

  1. Multi-perspective narrative interaction: CML allows users to engage in real-time dialogues with characters from different "universes" (e.g., movie characters and directors), breaking away from traditional one-way interaction models.
  2. Dynamic worldview boundaries: It explores cross-interactions ("collisions") among characters from different worldviews, encouraging users to evaluate narratives and themes from multiple perspectives.
  3. Personalized operation: Users can freely select agent combinations and design unique conversational paths.

• What are the implementation steps and key technologies used?

  1. Film recap zone: Users watch a 10-15 minute summary video of the film to enhance memory of the content.
  2. Interactive dialogue space: Users engage in conversations with virtual agents designed using the GPT-4o model. Key technologies include:
     • Character injection prompts: Each agent adopts specific prompt texts to embody its unique personality.
     • Dialogue coordination mechanism: An invisible coordinating agent manages the sequence and fluency of the conversation.
  3. Multi-agent selection mechanism: Users can choose from three agent groups—characters, creators, and virtual audiences—to match their conversational goals.

Research Outcomes

• What specific outcomes were achieved?

  1. Enhanced multi-perspective understanding: Users gained deeper insights into narratives and themes through interactions with different agents from internal (character perspectives) and external (creator and audience perspectives) viewpoints.
  2. Unexpected delightful experiences: Conflicts and collisions between cross-universe characters added novelty and fun, sparking users' enthusiasm for films.

• What advantages does it have compared to existing solutions?
  CML avoids the monotony of traditional one-on-one interactions by offering a dynamic discussion mechanism with multi-agents and multi-perspectives, enriching user engagement and interpretative depth.

• What were the experimental or evaluation results?

  • User study:
    • Participants engaged in three rounds of interactions with different agent combinations (e.g., protagonist, director, and virtual audience).
    • Results showed that 82% of users reported that the diverse perspectives provided by the agents helped them better understand the film's plot and themes.
  • Findings:
    • Multi-agent interactions reduced conversational pressure, encouraging free expression of opinions.
    • Users appreciated responses that offered associative or philosophical extensions based on character backgrounds (e.g., virtual characters discussing "cloning ethics").
    • For cross-universe discussions, users desired characters to maintain their own worldviews while presenting deeper conflicts.

• Limitations and Future Directions

  1. Experimental limitations: Participants did not directly watch the full film but reviewed its content through summary videos; future studies should attempt interactions immediately after full film viewing.
  2. Technical limitations: Current LLM models occasionally exhibit inconsistencies or silence when portraying characters. Future efforts could involve fine-tuning models with specialized datasets to enhance character consistency.
  3. Multi-cultural perspectives: The experiment was limited to a Korean cultural context; future research should include participants from diverse cultures to validate the universality of findings.
  4. Scalability: CML is currently suited for films with complex narratives and philosophical depth; further exploration is needed to adapt it to other genres (e.g., comedy or action films).

Through CML, the authors demonstrated the potential of multi-agent systems in enhancing film appreciation experiences and proposed important design principles for multi-agent interactive systems. This study lays a theoretical and practical foundation for future AI-based film media consumption models.