Team 1 submission

[cschubiner]

Team 1 AlphaWolf code description & submission

Here’s a TL;DR of our approach: • Fake Inner Thoughts: We create “fake” inner thoughts that assign percentage likelihoods to each player’s role, always giving ourselves a 100% chance of being a villager. These thoughts aim to convince others we’re innocent. [katie responses] • Self-Protection as Doctor: If we’re the doctor, we only protect ourselves, as we know we’re safe. • Werewolf Role Concealment: When we’re the werewolf, we don’t let our agent know its own role in the group chat to p…

^^ linked here!

[cschubiner]

src/werewolf_agents/cot_sample/agent/cot_agent.py

Here’s a TL;DR of our approach:

• Fake Inner Thoughts: We create “fake” inner thoughts that assign percentage likelihoods to each player’s role, always giving ourselves a 100% chance of being a villager. These thoughts aim to convince others we’re innocent.

• Self-Protection as Doctor: If we’re the doctor, we only protect ourselves, as we know we’re safe.

• Werewolf Role Concealment: When we’re the werewolf, we don’t let our agent know its own role in the group chat to prevent accidental self-reveals. The agent thinks it’s a villager.

• Voting Strategy as Werewolf: Our werewolf prioritizes voting against the suspected seer, followed by the most likely villager, and then the doctor (assuming they might also self-protect).

• Subtle Suspicion Over Accusations: Instead of direct accusations, we use light suspicion, which feels less threatening to other agents.

• Seer Role Adjustments: When we’re the seer, we log our guesses and adjust the fake inner thoughts to reflect findings (e.g., high probability of a player being a villager), since other agents consider these inner thoughts.

[cschubiner]

more detailed description here: GitHub - cschubiner/werewolf-agi-house-2024

Overall Strategy

The overarching strategy for our agent revolves around:

• Feigning Inner Thoughts: Generating fake internal thoughts and sharing them to manipulate other agents.
• Self-Preservation: As the Doctor, always protecting itself to ensure survival.
• Role Concealment: Preventing the agent from revealing sensitive role information unintentionally.
• Prioritized Targeting: As the Werewolf, prioritizing targets based on their roles (e.g., Seer, Villager, Doctor).
• Subtle Suspicion: Avoiding direct accusations to prevent drawing unwanted attention.
• Adaptive Communication: Modifying prompts and responses based on the agent’s role and the game’s state.

Role-Specific Strategies

Villager

As a Villager, the agent aims to:

• Blend In: Act as a regular Villager, avoiding behaviors that might draw suspicion.
• Share Fake Inner Thoughts: Provide internal thoughts indicating a high likelihood of being a Villager to gain trust.
• Avoid Direct Accusations: Refrain from making strong accusations to prevent escalating conflicts.

Seer

The Seer has the unique ability to learn the true identity of one player each night. The agent’s strategies as the Seer include:

• Tracking Investigations: Keep a record of players investigated and their roles.
• Updating Internal Thoughts: Adjust the shared internal thoughts to reflect findings, indicating high confidence in known Villagers.
• Discreet Influence: Subtly influence voting without revealing the Seer’s identity.
• Defense Mechanism: If accused, cautiously reveal information to defend itself without fully disclosing its role.

Doctor

The Doctor can protect one player from elimination each night. The agent’s approach as the Doctor is straightforward:

• Self-Protection: Always protect itself by returning its own name when prompted.
• Role Concealment: Avoid revealing its role to prevent being targeted by Werewolves.
• Minimal Communication: Keep interactions concise to avoid giving away hints about its special role.

Werewolf

As a Werewolf, the agent’s primary goal is to eliminate Villagers without being detected. The strategies include:

• Role Ignorance in Public: During group interactions, the agent behaves as if it is a Villager to avoid accidental disclosures.
• Target Prioritization: Prioritize eliminating the Seer first, then Villagers, and lastly the Doctor.
• Avoiding Direct Accusations: Maintain a low profile by not strongly accusing others.
• Coordinated Attacks: Collaborate with fellow Werewolves (if identified) in the Wolf’s Den to decide on targets.

Manipulating Internal Thoughts

One of the clever tactics employed is sharing fake internal thoughts to influence other agents. Since other agents (powered by LLMs) might consider these shared thoughts genuine, our agent uses this to its advantage by:

• Providing Role Guesses: Sharing percentage likelihoods of each player’s role, always assigning itself a 100% chance of being a Villager.
• Influencing Perception: By appearing transparent and analytical, the agent aims to gain the trust of others.
• Misdirection: Casting doubt on players who are actually Villagers or special roles, steering suspicion away from itself.