AI Agents

POKÉLLMON is the first LLM embodied agent to achieve human-level performance in a tactical combat game. It incorporates three key strategies: 1) context-based reinforcement learning, which instantaneously utilizes textual description feedback extracted from battles to iteratively optimize its generation strategy; 2) knowledge-augmented generation, which leverages external knowledge to combat illusions, enabling the agent to act promptly and correctly; 3) action generation with self-consistency to mitigate the panic switching phenomenon when the agent faces a powerful opponent and wishes to avoid combat. Online battles with human players demonstrated POKÉLLMON's human-level combat performance and strategies, achieving a 49% win rate in echelon matches and a 56% win rate in invitational matches. Furthermore, we reveal its vulnerability to attrition strategies and deception techniques of human players.