 |
DeepSeek is focusing on researching and developing new models, rather than appearing much in the media. Photo: SCMP . |
In collaboration with researchers from Tsinghua University, DeepSeek has introduced a new method to improve the reasoning ability of large language models (LLMs). The method, published in a research paper on the evening of April 4, helps LLMs produce better and faster results for common queries.
The technique is a combination of two of DeepSeek’s previous successful methods: generative reward modeling (GRM), which allows an AI model to self-evaluate and refine its answers based on previous results, and self-principled critique tuning.
Both methods rely on the “self-learning” element of AI, reducing the dependence on direct human feedback or guidance, but with the desire to produce results closer to human expectations.
Despite being a new method, DeepSeek-GRM still achieves outstanding performance and competes with the most famous and effective AI models today, according to the researchers. DeepSeek plans to open source the GRM models, but has not yet given a specific timeline.
After creating a global buzz with its V3 platform model and R1 inference model, DeepSeek published this academic paper on the online scientific repository arXiv, leaving many curious about the company's next move.
Reuters predicts that DeepSeek-R2, the successor to R1, will likely be launched in April, as previous models have yet to cool down. Previously, DeepSeek-R1 shocked the global technology world thanks to its outstanding performance compared to cost, enough to compete with current leading models.
DeepSeek has remained silent on the rumors. However, according to local information, a DeepSeek customer service account denied the information in a chat group with corporate customers.
Founded in Hangzhou in 2023 by entrepreneur Liang Wenfeng, DeepSeek has quickly attracted global attention in recent months. But rather than capitalize on its public fame, the company has focused its resources on research and development.
Previously, DeepSeek upgraded the V3 model, releasing version DeepSeek-V3-0324. According to the announcement, this update features enhanced inference capabilities, optimizations for web front-end development, and improved Chinese writing skills.
In February, the startup also open-sourced five code repositories, affirming its commitment to “advancement with full transparency.” That same month, it published a technical paper on “native sparse attention,” which improves the performance of LLMs when processing massive amounts of data.
DeepSeek is seen as a symbol of the resilience of China's AI industry, amid US efforts to curb the country's technological development.
Source: https://znews.vn/deepseek-gay-to-mo-post1543900.html
Comment (0)