Slicing as an approach to diffusion language model evaluation

A researcher describes a diffusion language model evaluation approach that uses extensive slicing to handle high dimensionality. They consider it more robust than earlier model-based approaches, but say it still reflects the biases of the model used as a critic; using GPT-2 to judge modern LLM outputs is their example. In an August essay, they also argue that evaluations lack standardization and can be biased by surrogate models.