Developer Creates an Interactive Fighting Game Arena for LLMs

A developer launched an online arena hosted at rawmodels.ai where users can pit different Large Language Models against each other in a fighting game. Fighters choose actions like moving, blocking, punching, or executing fireballs while attempting to manipulate opponents using conversational prompt injections. This project offers a creative intersection of game theory, strategy, and LLM adversarial testing, turning prompt manipulation into an engaging competitive sport. It provides a unique sandbox for testing how different models handle psychological and conversational attacks from opponents during runtime. Fighters can spend a turn sending a conversational attack into the opponent's prompt to influence their behavior, though it cannot alter the physical game state. The platform requires a registration fee of $1 USD, which covers about 100 fights, and tracks configurations with a rating and nickname system.