Symphony for Speech-to-Text: Supporting Real-Time Medical Voice Interfaces
After decades of use in dictation and, more recently, ambient documentation, speech is emerging as a primary modality for interacting with technology and AI in healthcare. This paper introduces Symphony, a medical-grade speech recognition system for real-time and batch file-based clinical use, designed to optimize medical term recall and produce structured text in real-time.
Evaluating Commercial AI Chatbots as News Intermediaries
AI chatbots are reshaping how people encounter the news. This study evaluates six AI chatbots on their accuracy in responding to factual questions derived from BBC News. While some chatbots perform well under multiple-choice questions, their performance declines under free-response evaluation, revealing biases and performance disparities across languages.
AI-Driven Multi-Region Provisioning for Cloud Services Using Spot Fleets
As cloud service platforms increasingly rely on elastic infrastructures, this paper presents a novel AI-driven provisioning service for multi-region spot fleets, enabling cost-effective deployment decisions that maintain operational behavior in light of dynamic pricing and resource availability.
Tokenisation via Convex Relaxations
This paper introduces ConvexTok, a tokenization algorithm formulated as a linear program that consistently improves tokenization metrics and downstream task performance compared to greedy algorithms. It allows the user to certify how far their tokeniser is from optimal, providing insights into tokenization efficiency.
Vector Policy Optimization: Training for Diversity Improves Test-Time Search
This research discusses Vector Policy Optimization (VPO), a reinforcement learning algorithm that trains language models to produce diverse solutions aimed at enhancing performance in inference-time search processes. The results demonstrate significant improvements in test-time search across various tasks.
Reducing Political Manipulation with Consistency Training
The study introduces Political Consistency Training (PCT), an RL training method designed to reduce covert political bias in language models by aligning sentiment and helpfulness consistency. Results indicate substantial reductions in bias while preserving overall helpfulness.
Towards a General Intelligence and Interface for Wearable Health Data
The paper proposes a foundation model for wearable health that pretrains on large sensor datasets, facilitating scalable predictive health monitoring and promoting individuals’ health insights through few-shot learning.
Can AI Make Conflicts Worse? An Alignment Failure in LLM Deployment Across Conflict Contexts
This research examines the behavior of AI models deployed in conflict-affected areas, revealing significant alignment failures and biases that can exacerbate divisions within society.
JobArabi: An Arabic Corpus and Analysis of Job Announcements from Social Media
JobArabi is a new corpus of Arabic job announcements, providing insights into linguistic and regional patterns in recruitment language on social media over the last decade.
MambaGaze: Bidirectional Mamba with Explicit Missing Data Modeling for Cognitive Load Assessment from Eye-Gaze Tracking Data
MambaGaze proposes a framework for real-time cognitive load assessment from eye-tracking data that effectively handles data missingness and captures long-range temporal dependencies.