Researchers Propose World-Time Compute Using Verified Code World Models

In a research preprint published on arXiv, scientists introduced an approach called 'world-time compute' to address data scarcity in large language model training. The method synthesizes executable and verified code-based world models that produce accurately labeled trajectories over symbolic states.

Fine-tuning language models on these verifiable environments improves generalization to unseen tasks, particularly benefiting smaller architectures. For instance, a 0.5B parameter model demonstrated a 29-point performance increase. The technique proved effective on benchmarks such as ARC-AGI, List Functions, and CLRS.

The environments are authored and executed through OpenWorld, a zero-dependency framework. The researchers noted that all code, recipes, and experimental workflows are fully reproducible from a single open repository.