Machine learning, particularly the training of large foundation models, relies heavily on the diversity and quality of data. These models, pre-trained on vast datasets, are the foundation of many modern AI applications, including language processing, image recognition, and more. The effectiveness of foundation models depends on how well they are trained, which is influenced by… →

Vision-Language Models (VLMs) are increasingly used for generating responses to queries about visual content. Despite their progress, they often suffer from a major issue: generating plausible but incorrect responses, also known as hallucinations. These hallucinations can lead to a lack of trust in these systems, especially in real-world, high-stakes applications. Evaluating the helpfulness and truthfulness… →

In machine learning, embeddings are widely used to represent data in a compressed, low-dimensional vector space. They capture the semantic relationships well for performing tasks such as text classification, sentiment analysis, etc. However, they struggle to capture the intricate relationships in complex hierarchical structures within the data. This leads to suboptimal performances and increased computational… →

Utilizing Large Language Models (LLMs) through different prompting strategies has become popular in recent years. However, many current methods frequently offer very general frameworks that neglect to handle the particular difficulties involved in creating compelling urges. Differentiating prompts in multi-turn interactions, which involve several exchanges between the user and model, is a crucial problem that… →

Multi-modal entity alignment (MMEA) is a technique that leverages information from various data sources or modalities to identify corresponding entities across multiple knowledge graphs. By combining information from text, structure, attributes, and external knowledge bases, MMEA can address the limitations of single-modal approaches and achieve higher accuracy, robustness, and effectiveness in entity alignment tasks. However,… →

Sparse autoencoders (SAEs) are an emerging method for breaking down language model activations into linear, interpretable features. However, they fail to fully explain model behavior, leaving “dark matter” or unexplained variance. The ultimate aim of mechanistic interpretability is to decode neural networks by mapping their internal features and circuits. SAEs learn sparse representations to reconstruct… →