Skip to content
NewA calmer, faster ZeroHour

GTAGym evaluates and trains agents on reconstructed GTA V tasks

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

Hugging Face daily papersby Jiajun Chen, Shicheng Fang, Jiahao Lu ()1 min read1
AI summary · x-ai/grok-4.7Summaries by Vani AI

GTAVBench has 103 online tasks in nine families; the best AGI setup scored 70.9%, while navigation and evasion stayed hard.

GTAGym is an environment for evaluating and reinforcement-learning agents on Grand Theft Auto V tasks. GTAVBench reconstructs 103 online main-story tasks across nine families and adds 47 image-based shooting and sniping tasks. A web server checks outcomes from engine state using screenshots and keyboard-mouse actions. Six frontier LLMs were tested, and LoRA RL on generated navigation tasks raised native GTA V performance from 15% to 23%.

Key facts
  • GTAVBench-AGI shows instructions only in the first screenshot.
  • A harness automatically compresses interaction history for that branch.
  • The generator samples task parameters under fixed templates for RL.
  • Qwen3.8-27B LoRA training transferred to out-of-domain task categories.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

1 min read at huggingface.co

AAA games offer a rich testbed for evaluating agent perception, reasoning, and control. However, evaluating whether agents can complete tasks in AAA games is challenging and often human-annotated because task outcomes are often difficult to verify automatically, let alone applying it for Reinforcement Learning (RL). We introduce GTAGym, an environment and framework for evaluation and RL on GTA V tasks.

Hugging Face daily papers

By Jiajun Chen, Shicheng Fang, Jiahao Lu · huggingface.co

Read the full article at Hugging Face daily papers

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.