The Cocktail Party

Overview

The Cocktail Party is a narrative-driven, social interaction game where players collaborate and compete to influence NPCs (Non-Player Characters) within a dynamically evolving scenario. Players are assigned unique roles, agendas, and items to create a rich, strategic gameplay experience.

The game includes the following key roles and components:

• Players: Attempt to persuade NPCs or spread/discredit specific narratives.
• NPCs: Communicative agents that propagate information and respond to players.
• Overseer: A game manager with omniscient knowledge of the scenario and player interactions.
• Scenarios: Narratives that provide the context for each game session.
• Items: Gameplay elements that introduce randomness and strategy.

Game Mechanics

Phases of Gameplay

1. Briefing Phase

   • Players receive context about the scenario.
   • Each player is assigned an agenda, a secret, and an item.
   • A set of "matters of fact" (MOF) is introduced, forming the backbone of the scenario.

2. Coercion Phase

   • Players interact with NPCs to influence them.
   • Players use their agendas, items, and knowledge of secrets to manipulate conversations.
   • NPCs have dynamic emotional states that evolve during the interaction.

3. Convening Phase

   • NPCs share their impressions with each other, forming a collective understanding.
   • The Overseer determines which players successfully influenced the NPCs and updates the game state.

4. Resolution Phase

   • Players receive feedback on their performance.
   • The scenario evolves based on player actions, leading to the next round or a conclusion.

5. Replay Phase

   • Players are offered the opportunity to replay the scenario or select a new one.
   • Additional insights or alternate paths can be explored based on previous decisions.

Player Mechanics

• Agendas: Players are given a specific narrative or goal they must push during interactions.
• Secrets: Each player possesses a unique secret tied to the scenario’s MOF. They can use this to gain trust or deceive others.
• Items: Items provide strategic advantages or obstacles. Examples include:
  • Truth Serum: Forces an NPC to reveal their secret.
  • Forged Evidence: Plants false information into the narrative.

NPC Mechanics

• Dynamic Emotional States:

  • NPCs have attributes such as:
    • Happiness
    • Hunger
    • Attentiveness
    • Defensiveness
    • Curiosity
  • These states evolve in response to player interactions, influencing their behavior.

• Memory System:

  • NPCs remember past interactions, forming biases toward players.

• Communication Propagation:

  • NPCs share information with each other, which may be distorted or misrepresented.

Overseer Mechanics

• Scenario Management:

  • Provides the scenario’s initial context and updates it dynamically as the game progresses.

• Game State Tracking:

  • Monitors player actions, NPC responses, and evolving narratives.
  • Determines if players achieve their agendas.

• NPC Behavior Management:

  • Oversees NPC emotional state changes and their propagation of information.

• Conversation Analysis:

  • Tracks whether players achieved their goals by analyzing dialogue and determining persuasion effectiveness.
  • Evaluates if key target statements were accepted or rejected by NPCs.

• Emotional Management:

  • Monitors the change in NPC dynamics, such as shifts in trust or defensiveness, as the game progresses.
  • Uses these metrics to influence the narrative and provide feedback to players.

Technical Design

Backend Infrastructure

The Cocktail Party uses a unified Intelligence Server as the central hub to manage gameplay logic and interactions. This server:

1. Monitors Active Game State:

   • Tracks players, NPCs, and narrative progress in real-time.

2. Creates Agent-Wise Conversations:

   • Generates contextually appropriate dialogues using AI models.
   • Adjusts NPC emotional states dynamically based on interactions.

3. Communicates with the Overseer:

   • Updates game mechanics, evaluates player influence, and triggers narrative changes.

4. Integrates with Unity Frontend:

   • Provides real-time feedback and dialogue for immersive player interactions.

Game Data

• Scenario Data:

  • Predefined narratives with associated MOF.

• Agent Data:

  • Includes NPC attributes, agendas, and emotional states.

• Interaction Data:

  • Tracks player-NPC interactions and updates NPC responses dynamically.

Fine-Tuning for Sentiment Analysis and Fallacy Detection

• The backend uses open-source models fine-tuned on datasets for sentiment analysis and logical fallacy detection.

• Open-Source Datasets: Examples include:

  • IMDB Sentiment Dataset for dialogue tone analysis.
  • ArguAna for logical consistency.

• Methods of Fine-Tuning:

  • Retrain models with contextual data from gameplay logs.
  • Incorporate domain-specific scenarios to align NPC responses with narrative themes.

• Fallacy Detection:

  • Identifies misleading arguments or logical inconsistencies in player-NPC conversations.
  • Flags conversations with high fallacy likelihood for Overseer review.

Fine-Tuning with Axolotl and Unsloth

To achieve optimal performance for in-game interactions, the backend leverages Axolotl and Unsloth for fine-tuning models:

1. Axolotl:

   • Provides an adaptable framework for fine-tuning large language models on game-specific datasets.
   • Used to train models to understand and replicate conversational dynamics, including nuanced NPC behaviors.
   • Supports integration with open-source datasets and custom gameplay logs to align with the game's narrative and interaction goals.

2. Unsloth:

   • Specializes in optimizing inference times and reducing latency during real-time gameplay.
   • Enhances the responsiveness of NPC interactions by pruning redundant model pathways while maintaining high fidelity in responses.

3. Workflow:

   • Data Preparation:
     • Gameplay logs and scenario-specific text are collected as training datasets.
     • Open-source data is preprocessed to fit narrative and interaction goals.
   • Model Training:
     • Axolotl is employed to fine-tune large language models to generate narrative-aligned conversations.
     • Unsloth optimizes these fine-tuned models for faster inference.
   • Evaluation:
     • Performance is measured against in-game metrics such as emotional state adaptation and fallacy detection accuracy.

Key Systems

• Dynamic Emotional State Generation:

  • Uses the Outlines library to generate values for NPC attributes (happiness, hunger, etc.).

• Scenario Evolution:

  • Scenario updates are triggered based on player actions and NPC feedback.

• Player-NPC Interaction Engine:

  • Processes player dialogue and generates NPC responses using LLaMA and Outlines.

Example Scenario

Murder Mystery on a Cruise

• Briefing Phase:

  • Context: A wealthy tycoon has been murdered aboard a cruise ship.
  • MOF: Key facts include the tycoon’s last known location, the suspects, and the timeline of events.
  • Player Agendas:
    • Frame another player for the murder.
    • Uncover the truth and exonerate an innocent suspect.

• Coercion Phase:

  • Players interact with NPCs, such as:
    • The ship’s captain (authoritative).
    • A passenger (unreliable witness).
    • The tycoon’s business rival (defensive).

• Convening Phase:

  • NPCs discuss their impressions and decide which narratives seem most credible.

• Resolution Phase:

  • The Overseer determines the outcome based on NPC decisions and updates the scenario.

• Replay Phase:

  • Players revisit the scenario with new insights or select an alternate path.

Future Improvements

• Expanded Scenarios:
  • Add more genres like political intrigue, corporate espionage, or alien diplomacy.
• Enhanced NPC AI:
  • Incorporate advanced memory and personality systems for NPCs.
• Player Progression:
  • Introduce unlockable items and scenarios based on player performance.