The world is being quietly rearranged by people who write very long documents.


The title they went with DiffScore: Text Evaluation Beyond Autoregressive Likelihood Noisy translates that to

AI models can now evaluate text without reading left-to-right


AI models that evaluate text usually read it from left to right, which can create a bias. Researchers have developed a new method, DiffScore, that reads text in all directions at once. This new method scores text quality more accurately on several benchmarks.
Current AI models often struggle to judge the overall quality of a piece of text because they process it sequentially. This new approach could lead to AI tools that better understand context and coherence. For now, this is a technical improvement in research, not a deployed change.
Watch for this method to be integrated into major open-source language model evaluation toolkits or adopted by companies building AI text generation products.

If you insist
Read the original →