After BrowseComp was brushed to 90%, Meituan LongCat launched LoHoSearch: Frontline models collectively dropped back to less than 30%

The strength of search agents has been largely determined by BrowseComp over the past year. However, this benchmark is losing its effectiveness — it pushed model performance from 30% to 90% in just 10 months, and the significance of the benchmark itself is rapidly being surpassed. On July 17, Meituan's LongCat introduced a tougher test called LoHoSearch, aiming to push search agents back into a phase of improvement.

LoHoSearch's challenges are not created by humans but are instead automatically generated from a knowledge graph of 7.62 million entities on Wikipedia. Due to their machine-generated origin, they reach difficulty levels that human annotators could never consistently design in terms of search space and structural complexity. In other words, previous benchmarks were designed by human brains, and even the hardest questions had a ceiling; LoHoSearch hands over the question-creation authority to the knowledge graph, completely removing the difficulty ceiling.

The results are striking. In tests with 11 cutting-edge models, the best score was only 34.74%, while the next three models scored around 15% to 16%. Meanwhile, the same top models currently achieve about 90% on BrowseComp. The gap clearly shows that LoHoSearch has exposed the true weaknesses of search agents. Even more interestingly, the marginal utility of context strategies: the best context strategy on LoHoSearch only brought a 6.8 percentage point improvement, while the same operation achieved 14 percentage points on BrowseComp — meaning that the old approach of compensating for ability gaps through prompts and context engineering is almost ineffective in this new benchmark.

This benchmark itself contains 544 questions covering 11 fields, organized in tree-like and graph structures, and has already been open-sourced. When the old benchmark reaches its peak, the real challenges for search agents have just begun, and LoHoSearch may become an unavoidable test item next.

Stack Overflow launches Stack Internal: Turning Enterprise Q&A into an AI Trustworthy Knowledge Base

Stack Overflow has launched the enterprise product Stack Internal, which provides technical Q&A metadata and reliability scores through the MCP interface, helping AI agents avoid generating incorrect information. The CEO revealed that large customers have already paid to use it, with a business model similar to Reddit's content licensing.

AI Math Learning Tool MathGPTPro: Supports Instant Answering Function for Uploaded Photo Problems

MathGPTPro is an advanced AI math learning tool that focuses on instant answering capabilities, supporting both image uploads and text inputs. It provides detailed solutions to math problems with a significantly higher accuracy than similar models. Its knowledge graph feature helps users track learning progress, identify strengths and areas for improvement, and achieve personalized learning. The product also includes real-time chat functionality and step-by-step guidance, making it easier to understand the problem-solving process and providing personalized learning suggestions to enhance the learning experience. It supports various input methods, including images, text, and LaTeX.

Goodbye Manual Labeling! AutoAlign Fully Automates Knowledge Graph Alignment with One Click

AutoAlign is an innovative knowledge graph alignment method designed to address the entity recognition challenges of different knowledge graphs. It constructs predicate proximity graphs using large language models to automatically identify similar predicates, achieving entity alignment between knowledge graphs. Without the need for manual seed alignment, it significantly enhances alignment efficiency and accuracy through predicate alignment, entity alignment, and a joint learning mechanism. In entity alignment tasks, AutoAlign outperforms existing methods and shows potential in applications such as knowledge graph completion.

Samsung Acquires UK AI Startup to Enhance Product Personalization Experience

Samsung Electronics has announced the acquisition of the UK-based AI startup Oxford Semantic Technologies to enhance its competitiveness in the field of artificial intelligence. Oxford Semantic Technologies is renowned for its knowledge graph technology, which is akin to human thinking, organizing data into interconnected networks for rapid information retrieval and personalized recommendations. The company's flagship product, RDFox, is hailed as one of the world's fastest knowledge graph and se

Meta Researchers Propose Lightweight Fine-tuning Method RA-DIT to Enhance Language Model Knowledge Retrieval Capabilities

RA-DIT improves the ability of language models to utilize retrieval information and provide relevant results from the retriever through a two-stage fine-tuning process. RA-DIT outperforms existing retrieval-augmented models in knowledge-intensive zero-shot and few-shot testing. The RA-DIT65B model shows improved performance in tasks that require knowledge, demonstrating the effectiveness of fine-tuning. The enhancement of retrieval capabilities brought by RA-DIT has shown good results in knowledge-intensive evaluations, confirming the effectiveness of fine-tuning.