The ability of learning to evaluate is increasingly taking on a pivotal role in the development of modern large multimodal models (LMMs). As pre-training on existing web data reaches its limits, researchers are shifting towards post-training with AI-enhanced synthetic data. This transition highlights the growing importance of learning to evaluate in modern LMMs. Reliable AI…
Transformers have gained significant attention due to their powerful capabilities in understanding and generating human-like text, making them suitable for various applications like language translation, summarization, and creative content generation. They operate based on an attention mechanism, which determines how much focus each token in a sequence should have on others to make informed predictions.…
Large Language Models (LLMs) have made considerable advancements in natural language understanding and generation through scalable pretraining and fine-tuning techniques. However, a major challenge persists in enhancing LLMs’ reasoning abilities, particularly for complex logical and mathematical tasks. The scarcity of high-quality preference data for fine-tuning reward models (RMs) limits the effectiveness of Reinforcement Learning from…
Introduction Traditional depth estimation methods often require metadata, such as camera intrinsics, or involve additional processing steps that limit their applicability in real-world scenarios. These limitations make it challenging to produce accurate depth maps efficiently, especially for diverse applications like augmented reality, virtual reality, and advanced image editing. To address these challenges, Apple introduced Depth…
Large language models (LLMs) have revolutionized natural language processing and artificial intelligence, enabling a variety of downstream tasks. However, most advanced models focus predominantly on English and a limited set of high-resource languages, leaving many European languages underrepresented. This lack of linguistic diversity creates significant barriers for non-English speakers, limiting their access to the capabilities…
In-context learning (ICL) enables LLMs to adapt to new tasks by including a few examples directly in the input without updating their parameters. However, selecting appropriate in-context examples (ICEs) is critical, especially for functions like math and logic that require multi-step reasoning. Traditional text-based embeddings often prioritize shallow semantic similarities, which may not align with…
Speech and audio processing is crucial in models involving speech data, particularly in handling complex tasks such as speech recognition, text-to-speech synthesis, speaker recognition, and speech enhancement. The key challenge lies in the variability and complexity of speech signals, which are influenced by factors like pronunciation, accent, background noise, and acoustic conditions. Additionally, the scarcity…
Language models have made significant strides in mathematical reasoning, with synthetic data playing a crucial role in their development. However, the field faces significant challenges due to the closed-source nature of the largest math datasets. This lack of transparency raises concerns about data leakage and erodes trust in benchmark results, as evidenced by performance drops…
Modern machine learning (ML) phenomena such as double descent and benign overfitting have challenged long-standing statistical intuitions, confusing many classically trained statisticians. These phenomena contradict fundamental principles taught in introductory data science courses, especially overfitting and the bias-variance tradeoff. The striking performance of highly overparameterized ML models trained to zero loss contradicts conventional wisdom about…