LLM Agents Play Pokemon Showdown
Summary
The article presents a study of multi-agent coordination between LLM-based players in a Pokémon Showdown 2v2 setting. It compares different talk modes (free talk, one message each, single agent controlling both, and no talk) using Qwen3.8-27B and quantifies win rates and token costs, showing that free dialogue improves collaboration but increases prompt tokens. The work aims to establish a benchmark for evaluating cooperative LLM behavior and potential RL training for better collaboration.