GTAGym: Real AAA Game Tasks Environment for Evaluation and Reinforcement Learning

GTAGym. Real AAA Game Tasks Environment for Evaluation and Reinforcement Learning. By Jiajun Chen and four others, from the Shanghai Innovation Institute and Fudan University. Published on the ninth of November.

AAA games make a rich testbed for agents, because playing one well asks for perception, reasoning and control at the same time. The trouble is checking the result. Whether an agent actually completed a task in such a game is often difficult to verify automatically, so evaluation has tended to fall back on human annotation. And if success cannot be checked by a machine, it cannot easily become a reward either, which puts reinforcement learning on these games further out of reach still.

GTAGym is an environment and framework for both evaluation and reinforcement learning on GTA V tasks. Its evaluation benchmark is called GTAVBench, and it is derived from the game's main-story missions. It holds one hundred and three online tasks, reconstructed across nine task families. Alongside them sit forty-seven image-based shooting and sniping tasks, built from GTA V imagery. What gets measured is whether an agent, seeing screenshots and acting through keyboard and mouse, carries a task through to completion.

The online environment and the evaluation interface are reached through a web server, which supports testing in parallel. Outcomes are checked automatically against the game engine's state, rather than by a person watching. There are two ways in. The main benchmark gives the agent its task as textual instructions. A branch called GTAVBench-AGI is stricter: the instructions appear only in the first screenshot, and a harness automatically compresses the interaction history as the agent goes.

For training, GTAGym also provides a scalable generator. Under fixed templates, it samples task parameters within ranges and option sets that the user specifies, and produces batches of 3D and image-based shooting environments. Each one comes with automatic reward and termination signals, so reinforcement learning can run without anyone scoring the outcomes by hand.

Six frontier LLMs were evaluated, both on GTAVBench with textual instructions and on the AGI branch. The best configuration reaches seventy point nine percent overall accuracy on GTAVBench-AGI. Some categories stay hard all the same: aerial navigation, evasion and interception remain difficult.

On the training side, Qwen3.8-27B was trained with LoRA reinforcement learning on generated on-foot navigation tasks. Its in-domain performance improves, and the improvement transfers to task categories outside that domain. On native GTA V tasks, the score rises from fifteen percent to twenty-three percent.

Two halves share one framework. One is a benchmark whose outcomes are judged by the engine instead of by annotators. The other is a generator that turns that same automatic checking into reward and termination signals for training. The headline accuracy shows frontier models getting a good way through story missions, while flying, evading and intercepting still mark where they fall short.