Chinese artificial intelligence startup DeepSeek and Tsinghua University have jointly developed a new model that can reportedly improve the reasoning ability of large language models, reduce the amount of training to reduce operating costs, and may be applied to the upcoming next-generation large model R2.
Combined with reports from Bloomberg and the South China Morning Post, DeepSeek and researchers at Tsinghua University have collaborated to detail in a paper published on Friday (April 4) a new reinforcement learning method that can improve the efficiency of artificial intelligence models.
They developed an approach that combines "Generalist Reward Modeling" (GRM) and "Self-Principled Critique Tuning," Make large language models answer general query questions better and faster.
In the paper, the researchers say the new approach helps AI models better follow human preferences and outperform existing methods and models on various benchmarks by rewarding more accurate and understandable responses. The results show that you can get better performance using fewer computing resources.
The paper calls the new model "DeepSeek-GRM" and it will be released as open source, but it doesn't give a specific timeline. The Massachusetts Institute of Technology Technology Review reports that the new training method may be applied to DeepSeek's next-generation large model R2.
DeepSeek's R1, a low-cost large model launched in January this year, attracted global attention. In February, Reuters quoted people familiar with the matter as saying that the company, eager to capitalize on its rising popularity, may bring forward the R2 model, which was originally scheduled to be released in May.
According to the latest analysis by industry experts at Cox Motors, it is expected that new and used car prices in the United States will significantly rise this year under President Donald Trump's 25% car tariffs.
According to the latest analysis by industry experts at Cox…
Recently, the European Commission Commissioner Sefjovic ann…
Recently, the news that the South Korean government plans t…
Recently, the news that the world's first sea land integrat…
Recently, the US stock market has experienced another epic …
In April 2025, the global trade landscape was thrown into t…