GTAGym: Real AAA Game Tasks Environment for Evaluation and Reinforcement Learning
GTAGym uses GTA V’s engine state to automatically check whether AI agents complete tasks.
Its benchmark includes 103 online tasks across nine mission-derived families, plus 47 image-based shooting and sniping tasks. Agents interact through screenshots and keyboard-mouse actions, with a web interface supporting parallel testing. Across six frontier LLMs, the best configuration reached 70.9% accuracy on GTAVBench-AGI, where instructions appear only in the first screenshot; aerial navigation, evasion and interception remained difficult. Reinforcement learning on generated on-foot navigation tasks also improved Qwen3.8-27B’s performance on native GTA V tasks from 15% to 23%.
HF Daily Papers' note
Its benchmark includes 103 online tasks across nine mission-derived families, plus 47 image-based shooting and sniping tasks. Agents interact through screenshots and keyboard-mouse actions, with a web interface supporting parallel testing. Across six frontier LLMs, the best configuration reached 70.9% accuracy on GTAVBench-AGI, where instructions appear only in the first screenshot; aerial navigation, evasion and interception remained difficult. Reinforcement learning on generated on-foot navigation tasks also improved Qwen3.8-27B’s performance on native GTA V tasks from 15% to 23%.
HF Daily Papers' note
score 5