We propose a retrieval-augmented LLM system that automatically identifies and rewrites problematic industrial contract revisions, achieving high accuracy even in low-resource legal settings.

We propose a communication-efficient distributed training framework (DPPF) that introduces a pushing force to promote wide minima recovery, improving generalization without incurring significant computational or communication overhead.

We analyze the relationship between model performance and training dataset size for end-to-end autonomous driving.

We introduce gradient-norm–weighted averaging methods for distributed deep learning that pull workers toward flatter regions of the loss landscape. The proposed model-level and layer-level variants converge in both convex and non-convex settings and empirically achieve faster training, better optima, and reduced communication compared to prior work.

We propose an automated, bi-level document classification pipeline for regulatory text that first filters irrelevant documents using an ensemble of classical classifiers and then routes relevant documents to the appropriate departments using a transformer-based DocBERT model, significantly improving efficiency and accuracy over single-model approaches when evaluated on a large real-world corpus from Con Edison.

We introduce ConFER, a continual-learning benchmark for facial expression recognition that evaluates how well popular CL methods adapt to real-world, incremental data without catastrophic forgetting, demonstrating state-of-the-art performance across multiple FER datasets and highlighting the benefits of CL for robust affective computing.