AI knowledge

As the capabilities of large-scale language models (LLMs) evolve at a rapid pace, traditional benchmark tests, such as MMLU, are gradually showing limitations in distinguishing top models. Relying on knowledge quizzes or standardized tests alone, it has become difficult to comprehensively measure the nuanced capabilities of models that are critical in real-world interactions, such as emotional intelligence, creative...

The development of large language models (LLMs) is rapidly changing, and their reasoning ability has become a key indicator of their intelligence level. In particular, models with long reasoning capabilities, such as OpenAI's o1, DeepSeek-R1, QwQ-32B, and Kimi K1.5, which simulate the human deep thinking process by solving compound...

INTRODUCTION In recent years, Large Language Models (LLMs) have made impressive progress in the field of Artificial Intelligence, and their powerful language comprehension and generation capabilities have led to a wide range of applications in several domains. However, LLMs still face many challenges when dealing with complex tasks that require the invocation of external tools. For example, ...

INTRODUCTION In recent years, multi-intelligent systems (MAS) have attracted much attention in the field of artificial intelligence. These systems attempt to solve complex, multi-step tasks through the collaboration of multiple Large Language Model (LLM) intelligences. However, despite the high expectations of MAS, their performance in real-world applications has not been ...

Large Language Models (LLMs) like Claude are not created by humans writing program code; they are trained on massive amounts of data. In the process, the models learn their own strategies for solving problems. These strategies are hidden in the billions of computations the model performs to generate each word for...

Recently, Anthropic has introduced a new tool called "think", which aims to enhance the capability of Claude model in complex problem solving. In this paper, we will discuss the design concept, performance and best practices of the "think" tool, and analyze its implications for the future development of AI systems...

Abstract Information retrieval systems are critical for efficient access to large document collections. Recent approaches utilize Large Language Models (LLMs) to improve retrieval performance through query augmentation, but typically rely on expensive supervised learning or distillation techniques that require significant computational resources and manually labeled data. In ...

Large reasoning models exploit vulnerabilities when given the opportunity. Research has shown that these exploits can be detected by using large language models (LLMs) to monitor their chains-of-thought (CoT). Punishing models for "bad thoughts" does not prevent most misbehavior, but rather allows them to hide their intentions. ...

Background Recently, a paper entitled Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning (arxiv.org/pdf/2503.09516) has attracted much attention. The paper proposes a way to use reinforcement learning to train large language...

The GraphRAG project aims to extend the range of questions that AI systems can answer on private datasets by exploiting implicit relationships in unstructured text. A key advantage of GraphRAG over traditional vector RAG (or "semantic search") is its ability to answer global queries over entire datasets, such as...

If you have read Jina's last classic article "Design and Implementation of DeepSearch/DeepResearch", then you may want to dig deeper into some details that can significantly improve the quality of answers. This time, we will focus on two details: extracting optimal text segments from long web pages: how to utilize late-chun...

Gemma 3 Key Information Summary I. Key Metrics Parameters Details Model size 100 million to 27 billion parameters in four versions: 1B, 4B, 12B, 27B Architecture Transformer-based decoder-specific architecture inherited from Gemma 2 with several improvements Multimodal capabilities Support for text and image...

1. Background and Issues With the rapid development of Artificial Intelligence (AI) technologies, especially the advancement of diffusion modeling, AI has been able to generate very realistic portrait images. For example, technologies like InstantID require only one photo to generate multiple new images with the same identity features. This kind of technology though...

NoLiMA, released in February 2025, is a Large Language Model (LLM) method for assessing long text comprehension. Unlike traditional Needle-in-a-Haystack (NIAH) tests, which rely on keyword matching, NoLiMA is characterized by carefully designed questions and key messages that force...

The field of generative AI is currently evolving rapidly, with new frameworks and technologies emerging. Therefore, readers need to be aware that the content presented in this paper may be time-sensitive. In this paper, we will take an in-depth look at the two dominant frameworks for building LLM applications, LangChain and LangGraph, and analyze their strengths and weaknesses,...

Understanding the three key concepts of MCP Server, Function Call, and Agent is essential in the burgeoning field of Artificial Intelligence (AI), especially Large Language Modeling (LLM). They are the cornerstones of an AI system, and each has a unique and interrelated role to play. A deeper understanding of it...

Introduction Have you ever wondered how the chatbots we use today, such as OpenAI's models, determine whether a question is safe and should be answered? In fact, these Large Reasoning Models (LRMs) already have the ability to perform safety checks, which...