An arXiv announcement introduces Grid-Bench, a benchmark for evaluating whether large language models can predict how discrete worlds change. The task centers on spatial instructions—moving objects, transforming regions, or modifying surroundings—and asks models to track how these operations alter a grid-like configuration.
According to the abstract, the benchmark emphasizes sequentiality: when multiple operations are applied one after another, each operation acts on a configuration already changed by previous steps. This makes the problem more than a single-step spatial reasoning test; a model must maintain an updated representation of the world as it changes.
The visible abstract stops before reporting results, so the announcement does not yet say how well LLMs perform on Grid-Bench. At this stage, the contribution is the benchmark itself and the question it poses.