Code generation models have made remarkable progress through increased computational power and improved training data quality. State-of-the-art models like Code-Llama, Qwen2.5-Coder, and DeepSeek-Coder show exceptional capabilities across various programming tasks. These models undergo pre-training and supervised fine-tuning (SFT) using extensive coding data from web sources. However, the application of reinforcement learning (RL) in code generation… →

The integration of visual and textual data in artificial intelligence presents a complex challenge. Traditional models often struggle to interpret structured visual documents such as tables, charts, infographics, and diagrams with precision. This limitation affects automated content extraction and comprehension, which are crucial for applications in data analysis, information retrieval, and decision-making. As organizations increasingly… →

After the success of large language models (LLMs), the current research extends beyond text-based understanding to multimodal reasoning tasks. These tasks integrate vision and language, which is essential for artificial general intelligence (AGI). Cognitive benchmarks such as PuzzleVQA and AlgoPuzzleVQA evaluate AI’s ability to process abstract visual information and algorithmic reasoning. Even with advancements, LLMs… →

Reinforcement learning (RL) for large language models (LLMs) has traditionally relied on outcome-based rewards, which provide feedback only on the final output. This sparsity of reward makes it challenging to train models that need multi-step reasoning, like those employed in mathematical problem-solving and programming. Additionally, credit assignment becomes ambiguous, as the model does not get… →

Aligning large language models (LLMs) with human values remains difficult due to unclear goals, weak training signals, and the complexity of human intent. Direct Alignment Algorithms (DAAs) offer a way to simplify this process by optimizing models directly without relying on reward modeling or reinforcement learning. These algorithms use different ranking methods, such as comparing… →

LLM inference is highly resource-intensive, requiring substantial memory and computational power. To address this, various model parallelism strategies distribute workloads across multiple GPUs, reducing memory constraints and speeding up inference. Tensor parallelism (TP) is a widely used technique that partitions weights and activations across GPUs, enabling them to process a single request collaboratively. Unlike data… →