GameWorld: Towards Standardized and Verifiable Evaluation of Multimodal Game Agents

2026-04-08

AI Summary

Existing multimodal game agents lack a standardized and verifiable evaluation framework, hindered by heterogeneous action interfaces and heuristic validation methods. To address this gap, this work proposes GameWorld, a benchmark comprising 34 browser-based games and 170 tasks that enables closed-loop evaluation of both computer-operation-based and semantic-action-based agents. The benchmark introduces a deterministic semantic action space, native keyboard-and-mouse control interfaces, verifiable state-based metrics, and high-fidelity browser simulation, thereby supporting standardized, reproducible, and outcome-oriented assessment. Experiments across 18 model–interface combinations demonstrate the robustness of the benchmark and reveal a substantial performance gap between current state-of-the-art agents and human-level proficiency.

Abstract

Towards an embodied generalist for real-world interaction, Multimodal Large Language Model (MLLM) agents still suffer from challenging latency, sparse feedback, and irreversible mistakes. Video games offer an ideal testbed with rich visual observations and closed-loop interaction, demanding fine-grained perception, long-horizon planning, and precise control. However, systematically evaluating these capabilities is currently hindered by heterogeneous action interfaces and heuristic verification. To this end, we introduce GameWorld, a benchmark designed for standardized and verifiable evaluation of MLLMs as generalist game agents in browser environments. Two game agent interfaces are studied: (i) computer-use agents that directly emit keyboard and mouse controls, and (ii) generalist multimodal agents that act in a semantic action space via deterministic Semantic Action Parsing. GameWorld contains 34 diverse games and 170 tasks, each paired with state-verifiable metrics for outcome-based evaluation. The results across 18 model-interface pairs suggest that even the best performing agent is far from achieving human capabilities on video games. Extensive experiments of repeated full-benchmark reruns demonstrate the robustness of the benchmark, while further studies on real-time interaction, context-memory sensitivity, and action validity expose more challenges ahead for game agents. Together, by offering a standardized, verifiable, and reproducible evaluation framework, GameWorld lays a robust foundation for advancing research on multimodal game agents and beyond. The project page is at https://gameworld-bench.github.io.

Problem

• multimodal game agents
• standardized evaluation
• verifiable benchmark
• action interfaces
• MLLM evaluation

Innovation

• GameWorld
• Multimodal Large Language Models
• Semantic Action Parsing
• Standardized Evaluation
• Verifiable Benchmark