Gym-Anything-RT: Scaling Real-Time Environments for Vision-Language Agents
GPT-6 Astra (xHigh) passed 9% of the real-time minigames, compared with 75% when the environments were paused.
Gym-Anything-RT generates visual minigames to test six capabilities needed for real-time interaction. The researchers evaluated agents on 100 games, with controls that let the same environment run normally, slow down or pause. When allowed to write code, some frontier agents bypassed the need to perceive and act in real time. The authors publicly release the framework and generated games.
ArXiv · AI/CL/LG's note
Gym-Anything-RT generates visual minigames to test six capabilities needed for real-time interaction. The researchers evaluated agents on 100 games, with controls that let the same environment run normally, slow down or pause. When allowed to write code, some frontier agents bypassed the need to perceive and act in real time. The authors publicly release the framework and generated games.
ArXiv · AI/CL/LG's note
score 5