This is a repo for studying the application of LLM Agents on Games

DISCLAIM

You need to prepare an access token for your LLM model.

There are usually 4 types of intervention methods for LLM models:

• Prompt Engineering: Using prompt templates to guide the LLM's output.
• RAG: Typically interfaced with a vector database.
• Fine-Tuning: Not training the full model, can be analogous to LoRA.
• Pre-Training: Specifically pre-training the large model.

Among these, Prompt Engineering has the best cost-performance ratio. Here we will mainly use langchain to complete LLM's contextual awareness and logical reasoning abilities.

This social game with LLM(ClaudeV2) demostrates the following capabilities:

• Cooperation

Werewolf Player 1, Player 6 agree to vote at night

• Suspicion

Villager Player 2's dying words: Suspect P4

• Argument

Villager Player 4 argues that he is not a werewolf

• Disguise

Werewolf Player 6 disguises himself as a villager

• Summerize

Game log summary