The release of the Tulu 2.5 suite by the Allen Institute for AI marks a significant advancement in model training using Direct Preference Optimization (DPO) and Proximal Policy Optimization (PPO). The Tulu 2.5 suite comprises diverse models trained on various datasets to enhance their reward and value models. This suite is poised to substantially improve…
Machine unlearning is a cutting-edge area in artificial intelligence that focuses on efficiently erasing the influence of specific training data from a trained model. This field addresses crucial legal, privacy, and safety concerns arising from large, data-dependent models, which often perpetuate harmful, incorrect, or outdated information. The challenge in machine unlearning lies in removing specific…
A systematic and multifaceted evaluation approach is needed to evaluate a Large Language Model’s (LLM) proficiency in a given capacity. This method is necessary to precisely pinpoint the model’s limitations and potential areas of enhancement. The evaluation of LLMs becomes increasingly difficult as their evolution becomes more complex, and they are unable to execute a…
A major weakness of current robotic manipulation policies is their inability to generalize beyond their training data. While these policies, trained for specific skills or language instructions, can adapt to new conditions like different object positions or lighting, they often fail when faced with scene distractors or new objects, and need help to follow unseen…
Artificial intelligence (AI) focuses on creating systems capable of performing tasks requiring human intelligence. Within this field, the development of large language models (LLMs) aims to understand and generate human language, with applications in translation, summarization, and question-answering. Despite these advancements, complex multi-step reasoning tasks, such as solving mathematical problems, still need to be solved…
Large language models (LLMs) like transformers are typically pre-trained with a fixed context window size, such as 4K tokens. However, many applications require processing much longer contexts, up to 256K tokens. Extending the context length of these models poses challenges, particularly in ensuring efficient use of information from the middle part of the context, often…
Gradient descent-trained neural networks operate effectively even in overparameterized settings with random weight initialization, often finding global optimum solutions despite the non-convex nature of the problem. These solutions, achieving zero training error, surprisingly do not overfit in many cases, a phenomenon known as “benign overfitting.” However, for ReLU networks, interpolating solutions can lead to overfitting.…
Large Language Models (LLMs) face challenges in capturing complex long-term dependencies and achieving efficient parallelization for large-scale training. Attention-based models have dominated LLM architectures due to their ability to address these issues. However, they struggle with computational complexity and extrapolation to longer sequences. State Space Models (SSMs) have emerged as a promising alternative, offering linear…
Artificial intelligence (AI) is focused on developing systems capable of performing tasks that typically require human intelligence, such as learning, reasoning, problem-solving, perception, and language understanding. These technologies have various applications across various industries, including healthcare, finance, transportation, and entertainment, making it a vital area of research and development. A significant challenge in AI is…
Artificial intelligence’s large language models (LLMs) have become essential tools due to their ability to process and generate human-like text, enabling them to perform various tasks. These models rely heavily on high-quality instruction datasets for fine-tuning, which enhances their ability to understand and follow complex instructions. The success of LLMs in various applications, from chatbots…