As AI systems become integral to daily life, ensuring the safety and reliability of LLMs in decision-making roles is crucial. While LLMs have shown impressive performance across various tasks, their ability to operate safely and cooperate effectively in multi-agent environments still needs to be explored. Cooperation is critical in scenarios where agents work together to… →

AI advancements have led to the incorporation of a large variety of datasets for multimodal models, allowing for a more comprehensive understanding of complex information and a substantial increase in accuracy. Leveraging their advantages, multimodal models find applications in healthcare, autonomous vehicles, speech recognition, etc. However, the large data requirement of these models has led… →

Generative models have emerged as great tools for synthesizing complex data and enabling sophisticated industry predictions. In recent years, their application has expanded beyond NLP and media generation to fields like finance, where the challenges of intricate data streams and real-time analysis demand innovative solutions. Generative foundation models thrive on three primary elements:  A large… →

Autoregressive models are used to generate sequences of discrete tokens. The next token is conditioned by the preceding tokens in a given sequence in the approach. Recent research showed that generating sequences of continuous embeddings autoregressively is also feasible. However, such Continuous Autoregressive Models (CAMs) generate these embeddings similarly sequentially, but they face challenges such… →

The field of natural language processing (NLP) has grown rapidly in recent years, creating a pressing need for better datasets to train large language models (LLMs). Multilingual models, in particular, require datasets that are not only large but also diverse and carefully curated to capture the nuances of many different languages. Existing resources like CC-100,… →

Web-crawled image-text datasets are critical for training vision-language models, enabling advancements in tasks such as image captioning and visual question answering.  However, these datasets often suffer from noise and low quality, with inconsistent associations between images and text that limit the capabilities of the models. This limitation prevents achieving strong and accurate results, particularly in… →

Code intelligence has grown rapidly, driven by advancements in large language models (LLMs). These models are increasingly utilized for automated programming tasks such as code generation, debugging, and testing. With capabilities spanning multiple languages and domains, LLMs have become crucial tools in advancing software development, data science, and computational problem-solving. The evolution of LLMs is… →

Large language models (LLMs) have profoundly influenced natural language processing (NLP), excelling in tasks like text generation and language understanding. However, the Arabic language—with its intricate morphology, varied dialects, and cultural richness—remains underrepresented. Many advanced LLMs are designed with English as their primary focus, leaving Arabic-centric models either overly large and computationally demanding or inadequate… →

Board games have long been pivotal in shaping AI, serving as structured environments for testing decision-making and strategy. Games like chess and Connect Four, with their distinct rules and varying levels of complexity, have enabled AI systems to learn dynamic problem-solving. The structured nature of these games challenges AI to anticipate moves, consider opponents’ strategies,… →

Reward modeling is critical in aligning LLMs with human preferences, particularly within the reinforcement learning from human feedback (RLHF) framework. Traditional reward models (RMs) assign scalar scores to evaluate how well LLM outputs align with human judgments, guiding optimization during training to improve response quality. However, these models often need more interpretability, are prone to… →