Multimodal large language models (MLLMs) focus on creating artificial intelligence (AI) systems that can interpret textual and visual data seamlessly. These models aim to bridge the gap between natural language understanding and visual comprehension, allowing machines to cohesively process various forms of input, from text documents to images. Understanding and reasoning across multiple modalities is…
Generative AI has emerged as a pivotal field with the rise of large language models (LLMs). These models are capable of producing complex outputs based on a variety of prompts. One notable area within this domain is Retrieval Augmented Generation (RAG), which integrates external information into LLMs to enhance factual accuracy. RAG specifically addresses the…
Efficient optimization of large-scale deep learning models remains a significant challenge as the cost of training large language models (LLMs) continues to escalate. As models grow larger, the computational burden and time required for training increase substantially, creating a demand for more efficient optimizers that can reduce both training time and resources. This challenge is…
Predicting the long-term behavior of chaotic systems, such as those used in climate modeling, is essential but requires significant computational resources due to the need for high-resolution spatiotemporal grids. One alternative to fully-resolved simulations (FRS) is to use coarse grids, with closure models correcting for errors by approximating the missing fine-scale information. While machine learning…
Previous research on reasoning frameworks in large language models (LLMs) has explored various approaches to enhance problem-solving capabilities. Chain-of-Thought (CoT) introduced articulated reasoning processes, while Tree-of-Thought (ToT) and Graph-of-Thought (GoT) expanded on this concept by incorporating branching possibilities and complex relationships between reasoning steps. Cumulative Reasoning (CR) introduced collaborative processes involving multiple specialized LLMs. These…
While LLMs have shown promise in natural language processing, they often need help to perform multi-step reasoning and problem-solving, particularly in areas that require abstract thinking and drawing inferences from incomplete or fragmented information. The ability to reason effectively is crucial for LLMs to be truly useful in real-world applications. This limitation hinders the application…
Neural networks are widely adopted in various fields due to their ability to model complex patterns and relationships. However, they face a critical vulnerability to adversarial attacks – small, malicious input changes that cause unpredictable outputs. This issue poses significant challenges to the reliability and security of machine learning models across various applications. While several…
In recent research, a state-of-the-art technique has been introduced for utilizing Large Language Models (LLMs) to verify RDF (Resource Description Framework) triples, emphasizing the significance of providing traceable and verifiable reasoning. The fundamental building blocks of knowledge graphs (KGs) are RDF triples, which are composed of subject-predicate-object statements that describe relationships or facts. Maintaining the…
LLMs exhibit remarkable language abilities, prompting questions about their memory mechanisms. Unlike humans, who use memory for daily tasks, LLMs’ “memory” is derived from input rather than stored externally. Research efforts have aimed to improve LLMs’ retention by extending context length and incorporating external memory systems. However, these methods do not fully clarify how memory…
Novak Zivanic has made a significant contribution to the field of Natural Language Processing with the release of Embedić, a suite of Serbian text embedding models. These models are specifically designed for Information Retrieval and Retrieval-Augmented Generation (RAG) tasks. Specifically, the smallest model in the suite has achieved a remarkable feat, surpassing the previous state-of-the-art…