The field of educational artificial intelligence is moving towards incorporating more nuanced and effective methods for assessing student performance and understanding. One notable direction is the development of techniques that can generalize to new and unseen assessment items, allowing for more flexible and accurate evaluations. Another area of focus is the improvement of automatic essay scoring, particularly in terms of assessing textual cohesion. Additionally, researchers are exploring new approaches to fine-tuning large pre-trained models, such as sparse fine-tuning, to adapt to specific tasks and domains. The use of sparse autoencoders is also being investigated for their ability to capture language-specific concepts and domain-specific features. Notable papers include:

• A novel item response theory approach to enhance essay cohesion assessment, which outperforms conventional machine learning models.
• A sparse fine-tuning framework for transformers that improves performance on generative tasks.
• A method for identifying language-specific features in large language models using sparse autoencoders.