Researchers have found that small language models can grade open-ended examination answers as accurately as more expensive models when they use an explicit grading rubric. The study, titled 'any-to-bench', tested the effectiveness of using a frontier model to extract questions and their corresponding rubrics from source documents, and then having lower-cost models perform repeated grading tasks. The results show that the accuracy of the grades depends heavily on the answer being graded, rather than the intelligence or effort of the grader. In fact, removing the official answer from the rubric causes a significant drop in reliability and leads to inflated scores. --- Why it matters: This study matters because it highlights the importance of using explicit grading rubrics in AI-assisted grading systems. By decoupling grading from judge intelligence, educators can ensure that grades are fair and consistent across different models and graders. Source: https://arxiv.org/abs/2608.17938