Search AI Products and News

Recently, a research team from Renmin University, Shanghai Artificial Intelligence Laboratory, University College London, and Dalian University of Technology revealed an important finding in the reasoning process of large models: when the model is thinking, the 'thinking words' it uses actually reflect a significant increase in its internal information. This research result provides a new perspective for better understanding the reasoning mechanisms of artificial intelligence through methods of information theory. You may have seen large models output some language that seems human-like when answering questions, such as "Hmm..." or "Let me think...".

In the field of artificial intelligence, reward models are a critical component for aligning large language models (LLMs) with human feedback, but existing models face the issue of "reward hacking." These models often focus on superficial features, such as the length or format of responses, rather than identifying genuine quality metrics, such as factual accuracy and relevance. The root cause lies in standard training objectives failing to distinguish between spurious associations and true causal drivers present in the training data. This failure leads to fragile reward models (RMs), which generate misaligned policies.

On July 4, 2025, Kunlun Xiwang continued to open-source the second-generation reward model Skywork-Reward-V2 series. This series includes 8 reward models based on different foundation models, with parameter sizes ranging from 600 million to 8 billion. Upon its release, it won all seven major reward model evaluation rankings, becoming a focus in the open-source reward model field. Reward models play a key role in the reinforcement learning from human feedback (RLHF) process. To build the next generation of reward models, Kunlun Xiwang has constructed a dataset containing 40 million

According to KPMG China's recent report, "The First 50 Health Tech Companies," China accounts for more than 70% of the global release volume of medical large models. This data not only demonstrates China's rapid development in the field of intelligent healthcare, but also reflects the wide application of large language models in the healthcare industry. The report points out that about 65% of the currently released medical large models are large language models. These models can process and generate natural language, playing a significant supporting role in the analysis of medical data, patient communication, and scientific research.

In the long-standing copyright infringement lawsuit filed by The New York Times against OpenAI, the case has made significant progress. According to Ars Technica, the federal judge presiding over the case has authorized The New York Times and its co-plaintiffs, The New York Daily News and the Investigative Reporting Center, to access OpenAI's user logs, including deleted content, to accurately determine the scope of the infringement. The New York Times believes that ChatGPT users may delete their history after bypassing the paywall, and therefore it is necessary to conduct large-scale data collection.

A recent report titled "Health Tech 50 - The First Edition" released by KPMG China reveals that China has taken a leading position in the field of medical large models globally. The report indicates that the number of medical large models launched in China accounts for more than 70% of the global total, far surpassing other countries and regions. In terms of model categories, large language models (LLMs) are the most numerous, accounting for nearly 65%. Moreover, the report also highlights the strong growth momentum of the intelligent medical devices market in China. It is expected that by 2025, the scale of the intelligent medical devices market in China will reach 24.23 billion yuan, and it will continue to grow.

Large Language Models (LLMs) have achieved significant progress in complex reasoning tasks by combining task prompts with large-scale reinforcement learning (RL), as demonstrated by models like Deepseek-R1-Zero, which directly apply reinforcement learning to base models, showcasing strong reasoning capabilities. However, this success is difficult to replicate across different base model families, especially within the Llama series. This raises a core question: what factors lead to inconsistent performance of different base models during reinforcement learning? How does reinforcement learning perform in

Baidu officially released the ERNIE Bot 4.5 series models and made them fully open source. Users can experience this latest open-source technology immediately through ERNIE Bot (https://yiyan.baidu.com). This series includes multiple parameter configurations, such as Mixture of Experts (MoE) models with activated parameters of 47B and 3B, as well as dense models designed with 0.3B parameters, totaling ten different models. In terms of training and inference, the ERNIE 4.5 series models use PaddlePaddle deep learning.

Baidu's WENXIN series large models have recently been open-sourced on its PaddlePaddle platform, including dozens of latest models such as ERNIE-4.5-VL-424B-A47B-Paddle and ERNIE-4.5-300B-A47B-Paddle. Although Baidu has not actively disclosed this open-source initiative, updates on the PaddlePaddle platform show that these actions were concentrated between June 29th and June 30th, marking its latest move. A source within the company confirmed: official

ByteDance's Seed Team recently released the impressive results of the 2025 full subject Gaokao test: the Dou Bao Seed1.6-Thinking model scored 683 in literature and 648 in science, meeting the admission threshold for Tsinghua and Peking University, and performed outstandingly in domestic and international AI model Gaokao tests. The test used the national new volume one and Shandong Province's independent proposition papers, with Dou Bao competing against five other top domestic and international AI models such as Google Gemini 2.5 Pro, DeepSeek R1, and OpenAI o3.