Game benchmarks often measure whether an agent reached a goal, but leave the hardest part—reliably determining success—to human reviewers. The key contribution here is to make GTA V missions both playable through a screenshot-and-action interface and verifiable from game state, connecting agent evaluation to repeatable reinforcement-learning experiments.
GTAGym: Real AAA Game Tasks Environment for Evaluation and Reinforcement Learning
Turns GTA V missions into automatically verifiable agent benchmarks and reinforcement-learning environments, with 150 navigation, interaction, and shooting tasks. Results expose persistent difficulty with dynamic control, while training on generated tasks improves some capabilities.
Introduction
Information
- Websitearxiv.org
- OrganizationsShanghai Innovation Institute, Fudan University
- AuthorsJiajun Chen, Shicheng Fang, Jiahao Lu, Yuxin Wang, Xipeng Qiu
- Published date2026/11/09
More Items
AI Agent Papers2026
Lets AI agents read task-relevant evidence from external content instead of receiving entire documents by default. Intent tracking, adaptive readers, and an exposure budget reduce token use and latency while retaining a faithful full-view fallback.
Trains language models to retrieve, study, and apply knowledge from textbooks through reinforcement learning. The approach raises overall mathematical proof accuracy from 51.5% to 79.5% and also improves results on two long-context benchmarks.
Lets a frozen model learn robot manipulation from one demonstration by rehearsing in a simulated world and storing feedback in context rather than weights. It reports 83.4% mean success across 50 RoboTwin 2.0 tasks.