The abstract of a new arXiv paper, "Words to World," argues that large language models are increasingly becoming an interface for action—invoking tools, operating devices, and making decisions that affect the physical world. However, the authors note, current benchmarks largely evaluate LLMs in ways that do not reflect these real-world interactions.
The paper proposes a new benchmark, "Words to World," designed to assess LLMs in interactive physical environments. While the abstract does not detail the benchmark's structure, its stated goal is to fill the gap left by existing evaluation methods, which tend to focus on language understanding and generation rather than embodied action.
The preprint is announced as new on arXiv, and the full details are not yet available from the abstract alone. The work highlights a growing need for benchmarks that measure not just what LLMs know, but what they can do in the world.