Open-source MLLMs exhibit considerable promise across diverse tasks by integrating visual encoders with language models. However, their reasoning abilities could be improved, largely due to existing instruction-tuning datasets often repurposed from academic resources like VQA and AI2D. These datasets focus on simplistic tasks with phrase-based answers and need more complexity for advanced reasoning. CoT reasoning,… →

DeepSeek AI has made significant progress in advancing artificial intelligence, particularly in areas like reasoning, mathematics, and coding. Earlier versions of its models achieved notable success in tackling mathematical and reasoning tasks, but there was room to improve their consistency across a broader range of applications, such as live coding and nuanced writing. These gaps… →

Neural networks (NNs) remarkably transform high-dimensional data into compact, lower-dimensional latent spaces. While researchers traditionally focus on model outputs like classification or generation, understanding the internal representation geometry has emerged as a critical area of investigation. These internal representations offer profound insights into neural network functionality, enabling researchers to repurpose learned features for downstream tasks… →

Transformers have been the foundation of large language models (LLMs), and recently, their application has expanded to search problems in graphs, a foundational domain in computational logic, planning, and AI. Graph search is integral to solving tasks requiring systematically exploring nodes and edges to find connections or paths. Despite transformers’ apparent adaptability, their ability to… →