A new arXiv paper argues that evaluation of large language models remains largely pre-scientific. The authors point to the many design decisions involved in benchmarking, and note that there is little guidance on how to make them. These choices, they say, can substantially affect measured performance.

The paper's title asks how to evaluate when you don't know how to evaluate, and frames the problem as one of choosing reliable setups for base language model benchmarking. Without a clear methodology, results may reflect arbitrary choices rather than model capability.

The authors suggest that the field needs more systematic approaches to benchmarking. Until then, comparisons between models should be treated with caution, as the setup itself may drive the outcome.