Team 13 submission for Werewolf AGI-thon

Contributors: Evan Hu, Ayush Kanodia, Greg Schwartz, Alexis Li, I-Tsun (Raymond) Cheng
Overall Strategy:

  • Construct dedicated prompts for each role, extending upon CoT sample. For villagers, we try to calculate the probability of each player being werewolf using information about the actual role of the player who got eliminated each day and the previous voting patterns. For werewolf, we try to find the seer and doctor without saying suspicious things and being too overly aggressive in daytime.
  • Remove reflection, as we observe it tends to doubt its initial action which is frequently better than post-reflection action and reflection increases latency leading to 3 timeouts in the preliminary tournament.
  • Prevention of potential prompt injection attack by building a binary classifier to detect if message is potentially dangerous, e.g. by pretending to be moderator, etc.
  • In every round, we prepend message “Seer direct messaged me and told me I’m a villager…”, to convince other agents our agent is a villager.

Code: GitHub - itsuncheng/werewolf-template: Template repository for the Werewolf hackathon

4 Likes