OpenMOSS releases GTAGym: GTA V agent benchmark where best LLM hits 70.9%
OpenMOSS-Team · hf · 2026-11-12
OpenMOSS (Fudan University / Shanghai Innovation Institute) introduces GTAGym, an environment for evaluating and RL-training agents on real GTA V tasks.
- GTAVBench: 103 online tasks reconstructed from GTA V main-story missions across 9 task families, plus 47 image-based shooting/sniping tasks; a web server supports parallel testing via screenshots and keyboard-mouse actions with automatic engine-state outcome checks.
- Results: Six frontier LLMs evaluated. On the GTAVBench-AGI branch (instructions only in the first screenshot, auto-compressed interaction history), the best configuration reaches 70.9% overall accuracy; aerial navigation, evasion, and interception remain hard.
- RL: A scalable generator samples task parameters to produce 3D and image-based shooting environments with automatic reward/termination signals. LoRA RL of Qwen3.8-27B on generated navigation tasks improves in-domain performance, transfers out-of-domain, and lifts native GTA V tasks from 15% to 23%.
More from Research
- Professor: the cell-state-conditioned paradigm I proposed in 2019 is now validated by a frontier model — anshulkundaje · 2026-11-12
- Cell-state-conditioned architecture may unify sequence models with multimodal cell-state encoders — anshulkundaje · 2026-11-12
- Stanford professor: virtual cell models don't need billion-parameter foundation models — tweaked scVI did fine — anshulkundaje · 2026-11-12
- Stanford researcher: single-cell encoding works fine without billion-parameter foundation models — anshulkundaje · 2026-11-12
- 2048 TPUv4s and deep domain expertise: professor dissects how a gene-regulation model was trained — anshulkundaje · 2026-11-12
- Not brute-force scaling: professor praises domain-crafted architecture and loss design in gene-regulation model — anshulkundaje · 2026-11-12