World Wire

Structure-from-motion (SfM) focuses on recovering camera positions and building 3D scenes from multiple images. This process is important for tasks like 3D reconstruction and novel view synthesis. A major challenge comes from processing large image collections efficiently while maintaining accuracy. Several approaches rely on the optimization of camera poses and scene geometry. However, these have usually increased computational costs substantially, and scaling SfM for large datasets remains challenging due to the sensitivity of balancing speed, accuracy, and memory consumption. Currently, SfM methods follow two main approaches: incremental and global. Incremental methods build 3D scenes step by step, starting from two images, while global methods align all cameras at once before reconstruction. Both rely on feature detection, matching, 3D triangulation, and optimization, leading to high computational costs and memory usage. Some learning-based methods improve accuracy but stru...

Large Language Models (LLMs) have become increasingly reliant on Reinforcement Learning from Human Feedback (RLHF) for fine-tuning across various applications, including code generation, mathematical reasoning, and dialogue assistance. However, a significant challenge has emerged in the form of reduced output diversity when using RLHF. Research has identified a critical trade-off between alignment quality and output diversity in RLHF-trained models. When these models align highly with desired objectives, they show limited output variability. This limitation poses concerns for creative open-ended tasks such as story generation, data synthesis, and red-teaming, where diverse outputs are essential for effective performance. Existing approaches to LLM alignment have focused on enhancing instruction following, safety, and reliability through RLHF, but these improvements often come at the cost of output diversity. Various methods have been developed to address this challenge, including the...

The rapid advancement of Large Language Models (LLMs) has significantly improved their ability to generate long-form responses. However, evaluating these responses efficiently and fairly remains a critical challenge. Traditionally, human evaluation has been the gold standard, but it is costly, time-consuming, and prone to bias. To mitigate these limitations, the LLM-as-a-Judge paradigm has emerged, leveraging LLMs themselves to act as evaluators. Despite this advancement, LLM-as-a-Judge models face two significant challenges: (1) a lack of human-annotated Chain-of-Thought (CoT) rationales, which are essential for structured and transparent evaluation, and (2) existing approaches that rely on rigid, hand-designed evaluation components, making them difficult to generalize across different tasks and domains. These constraints limit the accuracy and robustness of AI-based evaluation models. To overcome these issues, Meta AI has introduced EvalPlanner, a novel approach designed to improve...

Knowledge graphs have been used tremendously in the field of enterprise lately, with their applications realized in multiple data forms from legal persons to registered capital and shareholder’s details. Although graphs have high utility, they have been criticized for intricate text-based queries and manual exploration, which obstruct the extraction of pertinent information. With the massive strides in natural language processing and generative intelligence in the past years, LLMs have been used to perform complex queries and summarization based on their language comprehension and exploration skill set. This article discusses the latest research that uses language models to streamline information extraction from graph databases. Researchers from Baidu presented  “EICopilot,” an agent-based solution that streamlines search, exploration, and summarization of corporate data stored in knowledge graph databases to gain valuable insights about enterprises efficiently. To appreciate th...

Post-training quantization (PTQ) focuses on reducing the size and improving the speed of large language models (LLMs) to make them more practical for real-world use. Such models require large data volumes, but strongly skewed and highly heterogeneous data distribution during quantization presents considerable difficulties. This would inevitably expand the quantization range, making it, in most values, a less accurate expression and reducing general performance in model precision. While PTQ methods aim to address these issues, challenges remain in effectively distributing data across the entire quantization space, limiting the potential for optimization and hindering broader deployment in resource-constrained environments. Current Post-training quantization (PTQ) methods of large language models (LLMs) focus on weight-only and weight-activation quantization. Weight-only methods, such as GPTQ , AWQ , and OWQ , attempt to reduce memory usage by minimizing quantization errors or addres...

What is an Agent? An agent is a Large Language Model (LLM)-powered system that can decide its own workflow . Unlike traditional chatbots, which operate on a fixed path (ask → answer), agents are capable of: Choosing between different actions based on context. Using external tools such as web search, databases, or APIs. Looping between steps for better problem-solving. This flexibility makes agents powerful for complex tasks like conducting research, analyzing data, or executing multi-step workflows. Key Components of Agents Understanding the building blocks of agents is crucial before diving into implementation. These components work together to create intelligent, adaptable workflows: Agent (LLM Core) At the heart of every agent lies the “brain” of the system—the LLM. It is responsible for: Interpreting user inputs and understanding intent. Making decisions about the next steps based on pre-defined prompts and available tools. For example, when a user asks a questio...