Replay to Remember: Retaining Domain Knowledge in Streaming Language Models
This paper presents a method for continual learning in large language models (LLMs) focusing on addressing catastrophic forgetting through a lightweight approach combining LoRA and minimal replay mechanisms. The empirical results show significant stabilization of domain-specific knowledge, which is useful for real-world applications under resource constraints.
Conversational Assistants to support Heart Failure Patients: comparing a Neurosymbolic Architecture with ChatGPT
This paper evaluates two conversational assistants for heart failure patients, comparing a neurosymbolic architecture with ChatGPT. The study finds the in-house system outperforms ChatGPT in accuracy and task completion while also addressing speech errors, highlighting the practical challenges of different AI systems in healthcare.
Towards a HIPAA Compliant Agentic AI System in Healthcare
This work proposes a HIPAA-compliant AI framework for healthcare that addresses compliance through dynamic policy enforcement and provides mechanisms for managing protected health information, essential for integrating AI into clinical workflows.
Step1X-Edit: A Practical Framework for General Image Editing
This paper introduces Step1X-Edit, an open-source image editing model that combines multimodal LLM capabilities with user-driven instructions to establish benchmarks for real-world image editing applications, representing substantial advancements in the field.
EgoCHARM: Resource-Efficient Hierarchical Activity Recognition using an Egocentric IMU Sensor
EgoCHARM proposes a low-resource approach for human activity recognition using an egocentric IMU sensor, achieving impressive results with minimal hardware requirements, which is crucial for the development of efficient context-aware AI systems.
Cross-region Model Training with Communication-Computation Overlapping and Delay Compensation
This paper presents CoCoDC, an innovative framework for distributed training across geographic regions that optimizes communication overhead and computational staleness during model training, thereby enhancing efficiency and scalability.
The Role of Prescreening in Auctions with Predictions
The study develops a theoretical model exploring prescreening in auctions, showing that strategic prescreening can enhance revenue without the usual assumptions of bidder correlation, providing new insights into auction design.
Is Translation All You Need? A Study on Solving Multilingual Tasks with Large Language Models
This work evaluates whether translating queries into English truly enhances performance on multilingual tasks, revealing that native language prompts can more effectively capture cultural nuances, suggesting a re-evaluation of multilingual assessment methodologies.
The Sparse Frontier: Sparse Attention Trade-offs in Transformer LLMs
This research examines the trade-offs between efficiency and accuracy in sparse attention models, providing scaling laws and optimal strategies for using sparse attention in long-context processing, which is critical for advancements in transformer-based systems.