AI models can now be tested on specific skills, not just overall scores

Researchers have developed a new way to test large language models by breaking down their abilities into many specific skills, instead of giving them a single overall score. This means developers can now see exactly which skills an AI model is good at or bad at, making it easier to improve them or pick the right one for a job.

For years, evaluating AI models was like grading a student on a single test score, without knowing if they aced algebra but failed geometry. This new method provides a detailed report card, showing strengths and weaknesses across dozens of specific abilities in subjects like math, physics, and chemistry. This shift means AI developers can now target training to fix specific skill gaps, rather than guessing what needs improvement, and users can select models based on the exact skills required for a task.