GameWorld benchmark introduces 34 diverse games with 170 tasks to standardize evaluation of multimodal AI agents in realistic interactive environments.

GameWorld provides a standardized testing framework for evaluating Multimodal Large Language Model (MLLM) agents in browser-based game environments with rich visual feedback and closed-loop interaction.

Two agent interfaces are studied: computer-use agents that emit keyboard and mouse controls directly, and generalist multimodal agents that operate through semantic action spaces via deterministic Semantic Action Parsing.

The benchmark contains 34 diverse games with 170 paired tasks designed to test fine-grained perception, long-horizon planning, and precise control capabilities.

GameWorld addresses current limitations in MLLM agent evaluation including challenging latency, sparse feedback, irreversible mistakes, and heterogeneous action interfaces across different systems.