Generalist Foundation Models Are Not Clinical Enough for Hospital Operations

Hospitals and healthcare systems rely on operational decisions that determine patient flow, cost, and quality of care. Despite strong performance on medical knowledge and conversational benchmarks, foundation models trained on general text may lack the specialized knowledge required for these operational decisions. We introduce Lang1, a family of models (100M-7B parameters) pretrained on a specialized corpus blending 80B clinical tokens from NYU Langone Health's EHRs and 627B tokens from the internet. To rigorously evaluate Lang1 in real-world settings, we developed the REalistic Medical Evaluation (ReMedE), a benchmark derived from 668,331 EHR notes that evaluates five critical tasks: 30-day readmission prediction, 30-day mortality prediction, length of stay, comorbidity coding, and predicting insurance claims denial. In zero-shot settings, both general-purpose and specialized models underperform on four of five tasks (36.6%-71.7% AUROC), with mortality prediction being an exception. After finetuning, Lang1-1B outperforms finetuned generalist models up to 70x larger and zero-shot models up to 671x larger, improving AUROC by 3.64%-6.75% and 1.66%-23.66% respectively. We also observed cross-task scaling with joint finetuning on multiple tasks leading to improvement on other tasks. Lang1-1B effectively transfers to out-of-distribution settings, including other clinical tasks and an external health system. Our findings suggest that predictive capabilities for hospital operations require explicit supervised finetuning, and that this finetuning process is made more efficient by in-domain pretraining on EHR. Our findings support the emerging view that specialized LLMs can compete with generalist models in specialized tasks, and show that effective healthcare systems AI requires the combination of in-domain pretraining, supervised finetuning, and real-world evaluation beyond proxy benchmarks.

Refining Packing and Shuffling Strategies for Enhanced Performance in Generative Language Models

Packing and shuffling tokens is a common practice in training auto-regressive language models (LMs) to prevent overfitting and improve efficiency. Typically documents are concatenated to chunks of maximum sequence length (MSL) and then shuffled. However setting the atom size, the length for each data chunk accompanied by random shuffling, to MSL may lead to contextual incoherence due to tokens from different documents being packed into the same chunk. An alternative approach is to utilize padding, another common data packing strategy, to avoid contextual incoherence by only including one document in each shuffled chunk. To optimize both packing strategies (concatenation vs padding), we investigated the optimal atom size for shuffling and compared their performance and efficiency. We found that matching atom size to MSL optimizes performance for both packing methods (concatenation and padding), and padding yields lower final perplexity (higher performance) than concatenation at the cost of more training steps and lower compute efficiency. This trade-off informs the choice of packing methods in training language models.

3D U-KAN Implementation for Multi-modal MRI Brain Tumor Segmentation

We explore the application of U-KAN, a U-Net based network enhanced with Kolmogorov-Arnold Network (KAN) layers, for 3D brain tumor segmentation using multi-modal MRI data. We adapt the original 2D U-KAN model to the 3D task, and introduce a variant called UKAN-SE, which incorporates Squeeze-and-Excitation modules for global attention. We compare the performance of U-KAN and UKAN-SE against existing methods such as U-Net, Attention U-Net, and Swin UNETR, using the BraTS 2024 dataset. Our results show that U-KAN and UKAN-SE, with approximately 10.6 million parameters, achieve exceptional efficiency, requiring only about 1/4 of the training time of U-Net and Attention U-Net, and 1/6 that of Swin UNETR, while surpassing these models across most evaluation metrics. Notably, UKAN-SE slightly outperforms U-KAN.

Persona-centric Metamorphic Relation guided Robustness Evaluation for Multi-turn Dialogue Modelling

Recently there has been significant progress in the field of dialogue system thanks to the introduction of training paradigms such as fine-tune and prompt learning. Persona can function as the prior knowledge for maintaining the personality consistency of dialogue systems, which makes it perform well on accuracy. Nonetheless, the conventional reference-based evaluation method falls short in capturing the genuine text comprehension prowess of the model, significantly relying on the quality of data annotation. In contrast, the application of metamorphic testing offers a more profound insight into the model's distinct capabilities without necessitating supplementary annotation labels. This approach furnishes a more comprehensive portrayal of the model's intricacies and exposes intricacies concealed within reference-based validation techniques. Consequently, we introduce a persona-centric metamorphic relation construction for metamorphic testing, aimed at evaluating both the persona consistency and robustness of personalized dialogue models. For that reason, this work evaluates several widely used training paradigms including learning from scratch, pretrain + fine-tune and prompt learning in personalized dialogue retrieval to know if they are more robust or if they have the same flaws as their predecessor. Under three kinds of designed metamorphic relations with consistent outputs, our experimental results reveal that prompt learning shows stronger robustness compared to training from scratch and fine-tune. Although tested retrieval models gain competitively high retrieval accuracy according to the traditional reference-based validation, they are still fragile and demonstrate various unexpected behaviors, thus there is still room for future improvement in personalized dialogue retrieval.