On December 12, Peking University and ByteDance officially announced the establishment of the Doubao Large Model System Software Joint Laboratory, marking a new milestone in the collaboration between industry, academia, and research in China's AI field. The research work of the joint laboratory will be based on the Doubao large model independently developed by ByteDance. Researchers from universities and enterprises will conduct in-depth research on key technologies such as model training and inference within the laboratory. With the construction and development of the laboratory, a large number of Peking University students will also enter the laboratory for internships, working side by side with industry researchers.

Recently, research teams from Peking University announced the release of an open-source multimodal model called LLaVA-o1, which is claimed to be the first visual language model capable of spontaneous and systematic reasoning, comparable to GPT-o1. The model excels in six challenging multimodal benchmark tests, with its 11B parameter version outperforming competitors such as Gemini-1.5-pro, GPT-4o-mini, and Llama-3.2-90B-Vision-Instruct.

Kuaishou, Peking University, and Beijing University of Posts and Telecommunications have joined forces to unveil a groundbreaking technological achievement—the Pyramid-Flow high-definition video model. This open-source project has made significant breakthroughs in the field of AI-generated video, bringing new possibilities to the industry. The Pyramid-Flow model demonstrates remarkable capabilities, generating high-quality videos of up to 10 seconds in length, with a resolution of 1280x768 and a frame rate of 24fps, using only text input. It excels in light and shadow effects, motion continuity, overall image quality, and text interpretation.

Recently, Kuaishou, along with research teams from Peking University and Beijing University of Posts and Telecommunications, jointly open-sourced an ultra high definition video generation model named Pyramid-Flow. This model can generate videos of up to 10 seconds long, with a resolution of 1280x768 and 24 frames per second, producing high-quality results with excellent lighting effects, action consistency, and overall video quality. Pyramid-Flow operates differently from existing video diffusion models, which typically run at full resolution; while this approach allows for...

Following OpenAI's GPT-4 achieving remarkable results in traditional mathematics assessments, the research teams from Peking University and Alibaba have jointly launched a brand new evaluation benchmark - Omni-MATH, aimed at assessing the reasoning abilities of large language models at the level of the Olympic mathematics competitions. This initiative not only provides a new standard for evaluating AI mathematics capabilities but also opens up new avenues for exploring AI's potential in advanced mathematics. The unique design of Omni-MATH includes a database of 4,428 competition-level mathematics problems.

In the context of the rapid development of AI technology, prompt engineering has become a critical skill. By cleverly designing prompts, users can guide AI to better understand human intentions and improve the accuracy of AI responses. However, this process poses significant challenges for average users. To address this issue, a team of scientists from Peking University and Baichuan Technology has jointly developed the PAS (Prompt Augmentation System), an intelligent prompting enhancement system based on large language models. PAS improves data utilization, compatibility, and flexibility.

The research team led by Yuan Li at Peking University, in collaboration with researchers from Nanyang Technological University and Tsinghua University, has developed an innovative image retrieval technology that aims to precisely match user needs through sketches, artworks, and even low-resolution images. The core of this technology is the 'Universal Style Retrieval' method, which can handle multiple query styles including combinations of text, sketches, and artworks, significantly enhancing retrieval flexibility and accuracy. The research team constructed the DSR and ImageNet-X datasets to provide training and testing resources for the new method.

Long context understanding is a key challenge in the field of natural language processing, especially when large language models (LLMs) handle text that exceeds their context window size. To address this issue, researchers developed the LooGLE benchmark test, aimed at assessing LLMs' capability to understand extremely long documents (averaging 19.3k words, comprising 776 articles across multiple domains). LooGLE includes 7 tasks that cover both short and long dependencies, evaluating model understanding across texts of varying lengths. The test data is sourced from open-access content after 2022.

最近，来自北京大学和鹏城实验室的研究团队推出了一款名为 HoloDreamer 的新型3D 场景生成框架，旨在通过用户的文本描述生成身临其境、完整的3D 场景。

The team from Peking University and the Hong Kong University of Science and Technology made a big splash with a new training method that has achieved GPT-4 level performance with an 8B-sized medical expert model. This is no small feat, and they have also introduced a new concept, "stability gap," to explain certain phenomena observed during the continuous pre-training of large language models.Image Source Note: The image is generated by AI, and the image is provided by Midjourney, an image autho

Recently, Zhihua, Tsinghua University, and Peking University collaborated to launch a long video understanding benchmark project called LVBench. While existing multimodal large language models have made significant progress in understanding short videos, they still face challenges when dealing with long videos that can last for hours. LVBench was created to fill this gap.

The AI team from Peking University and Kuaishou has successfully tackled the challenge of complex video generation. They have proposed a new framework called VideoTetris, which can easily piece together various details like a puzzle to generate high-difficulty, complex-instruction videos. This framework surpasses commercial models such as Pika and Gen-2 in complex video generation tasks.

The significant effects of 90s parents from Tsinghua and Peking University using AI to comfort their children. The AI responses resonate with parents.

MegaScale has built a single cluster with over 10,000 GPUs, achieving a model FLOP utilization rate of 55.2%. MegaScale includes diagnostic tools for monitoring system components and events.

Is it possible to have a universal graph model that can predict toxicity based on molecular structure while also providing friend recommendations in social networks? Or one that can predict the citations of different authors to uncover human aging mechanisms in gene networks? It turns out, this model was jointly proposed by Professor Chen Yixin's team from Washington University in St. Louis, Zhang Muhan from Peking University, and researchers such as Tao Dacheng from JD.com Research Institute.

Peking University has introduced the GPT-4 based AI teaching assistant Brainiac Buddy to enhance learning and teaching experiences. Founder Dong Bin, with a background in mathematics and machine learning, aims to improve teaching efficiency. Students can hand over complex materials to the AI assistant, which highlights key content and answers questions, enhancing learning efficiency. Teachers can share lesson plans with the AI to assist in developing teaching strategies, saving time. Brainiac Buddy offers personalized and customized learning experiences, driving progress in the field of education.