Researchers have developed GTAGym, a new environment and framework designed to evaluate and train AI agents using tasks from the AAA game Grand Theft Auto V. The system includes GTAVBench, a benchmark with 103 main-story mission tasks and 47 shooting tasks, and GTAVBench-AGI for instruction-based evaluation. Initial tests show that while frontier LLMs achieve 70.9% accuracy on GTAVBench-AGI, complex tasks like aerial navigation remain challenging. The framework also supports reinforcement learning by generating task parameters with automatic reward signals, demonstrating improved performance for a Qwen3 model on navigation tasks. AI
IMPACT Provides a new, complex environment for testing and training AI agents, potentially advancing capabilities in perception, reasoning, and control.
RANK_REASON This is a research paper detailing a new environment and benchmark for AI agent evaluation. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
