Using agentic game development to scale spatial world models

Training large artificial intelligence models usually relies on gathering more crawled video data and throwing more compute power at the problem. A group of researchers argues that this approach is inefficient. Just as language models advanced rapidly because code execution environments provide concrete rewards, spatial world models need a similar foundation to move past blurry, biased metrics.

The limits of current spatial training

When training models that generate visual or spatial environments, developers often rely on fuzzy metrics like CLIP scores to measure quality. These alignment scores evaluate how well an image matches a text description, but they remain prone to bias and lack precision. Without exact feedback, it becomes very difficult to run reinforcement learning post-training — a process where a model refines its behavior through trial and error based on clear rewards.

Coding models succeed because computer code is executable. Compilers and software runtimes immediately tell an agent whether its code works or fails. Spatial generation currently lacks an equivalent environment that can offer clear, automated feedback for spatial reasoning.