Artificial intelligence research has steadily advanced toward creating systems capable of complex reasoning. Multimodal large language models (MLLMs) represent a significant development in this journey, combining the ability to process text and visual data. These systems can address intricate challenges like mathematical problems or reasoning through diagrams. By enabling AI to bridge the gap between…
The generation of synthetic tabular data has become increasingly crucial in fields like healthcare and financial services, where privacy concerns often restrict the use of real-world data. While autoregressive transformers, masked transformers, and diffusion models with transformers, have shown significant success in generating high-quality synthetic data with strong fidelity, utility, and privacy guarantees, they face…
Microsoft has open-sourced Phi-4, a compact and efficient small language model, on Hugging Face under the MIT license. This decision highlights a shift towards transparency and collaboration in the AI community, offering developers and researchers new opportunities. What Is Microsoft Phi-4? Phi-4 is a 14-billion-parameter language model developed with a focus on data quality and…
Language-based agentic systems represent a breakthrough in artificial intelligence, allowing for the automation of tasks such as question-answering, programming, and advanced problem-solving. These systems, heavily reliant on Large Language Models (LLMs), communicate using natural language. This innovative design reduces the engineering complexity of individual components and enables seamless interaction between them, paving the way for…
The o1 model’s impressive performance in complex reasoning highlights the potential of test-time computing scaling, which enhances System-2 thinking by allocating greater computational effort during inference. While deep learning’s scaling effects have driven advancements in AI, particularly in LLMs like GPT, further scaling during training faces limitations due to data scarcity and computational constraints. Additionally,…
The pre-training of language models (LMs) plays a crucial role in enabling their ability to understand and generate text. However, a significant challenge lies in effectively leveraging the diversity of training corpora, which often include data from varied sources such as Wikipedia, blogs, and social media. Models typically treat all input data equivalently, disregarding contextual…
Complex domains like social media, molecular biology, and recommendation systems have graph-structured data that consists of nodes, edges, and their respective features. These nodes and edges do not have a structured relationship, so addressing them using graph neural networks (GNNs) is essential. However, GNNs rely on labeled data, which is difficult and expensive to obtain.…
Large language models (LLMs) have revolutionized natural language processing, enabling applications that range from automated writing to complex decision-making aids. However, ensuring these models produce factually accurate responses remains a significant challenge. At times, LLMs generate outputs that appear credible but are factually incorrect, a phenomenon often referred to as “hallucination.” This issue becomes particularly…