POKÉLLMON: A Human-Parity Agent for Pokémon Battle with Large Language Models

Anonymous Authors

PokéLLMon battles against human players from the Ladder competitions.

The agent repeatedly uses an attack move that has zero effect to the opposing Pokémon before adopting the in-context reinforcement learning strategy

Due to the in-context reinforcement learning strategy, the agent changes is action after observing that the attack move has no effect to the opposing Pokémon

While facing a powerful opponent, the agent with the chain-of-thought reasoning panics and starts to switch Pokémon in consecutive turns to avoid battle.

An experienced human player misdirects PokéLLMon to waste its enhanced attack chance by first sending out a dragon-type pokémon and immediately switch to another Pokémon immune to the dragon-type attack.

PokéLLMon exhibits a human-like attrition strategy by poisoning the opposing Pokémon and frequently recovering its HP.

PokéLLMon is vulnerable to the human player's attrition strategy. The key to break this is to first boost its Pokemon's attack to a high stage and then cause unrecoverable damage.

Abstract

We introduce PokéLLMon, the first LLM-embodied agent that achieves human-parity performance in tactic battle games. It incorporates three key strategies: 1) In-context reinforcement learning that consumes text described feedback instantly derived from battles to iteratively refine its generation policy; 2) Knowledge-augmented generation that employs external knowledge to counteract hallucination and enables the agent to act timely and properly; 3) Action generation with self-consistency to mitigate the panic switching phenomenon when the agent faces a powerful opponent and want to avoid the battle. Online battle against human players demonstrate PokéLLMon's human-level battle performance and strategies, achieving 49% of wining rate in the ladder competitions and 56% of wining rate in the invited battles. In addition, we unveil its vulnerabilities to human players' attrition strategies and deception tricks.