A new arXiv paper investigates how pretrained language models can be adapted for univariate time-series forecasting using parameter-efficient transfer learning. The authors focus on identifying which design decisions matter most for effective transfer between text and numerical sequences. The work is a cross-listed submission to arXiv's machine learning category.
A new arXiv paper proposes knowledge-enriched structured features drawn from electronic health records to predict whether a patient will be readmitted within 30 days. The authors position the approach as an alternative to models that depend on discharge summaries and pretrained language models, which require clinical notes and carry heavy computational costs. Evaluation is conducted on the MIMIC-IV dataset.
A new arXiv paper revisits comparisons between state-of-the-art coreference resolution systems. Because every leading system fine-tunes a pretrained language model, the authors ask whether differences in scores come from the underlying language model or from task-specific design choices. The work presents a controlled reevaluation to separate those factors.