Chinese artificial intelligence (AI) start-up DeepSeek Has presented an innovative method for enhancing the reasoning abilities of large language models (LLMs) ahead of the anticipated launch of the firm's upcoming version.
DeepSeek, working alongside researchers from Tsinghua University, has devised a method that integrates techniques known as generative reward modeling (GRM) and self-guided critical fine-tuning, as detailed in a research paper released on Friday. This combined strategy seeks to enhance the efficiency and accuracy of responses provided by large language models when addressing broad inquiries.
The DeepSeek-GRM models produced better outcomes than current techniques, as they "attained comparable results" to leading public reward models, according to the research team. This method of reward modeling helps steer an LLM toward aligning with human preferences.
Are you looking for insights into the most significant issues and developments globally? Find your answers here. SCMP Knowledge Our latest platform features handpicked content including explainers, FAQs, analyses, and infographics, all provided by our esteemed team of experts.
DeepSeek planned to release the GRM models as open-source, stated the researchers; however, they didn't provide a specific timetable.
The research paper, available on the online scientific archive arXiv, emerges as speculations circulate regarding the startup's subsequent steps after the widespread interest sparked by their V3 foundational model and R1 reasoning model.
Last month, Reuters stated that DeepSeek-R2, which follows up on R1, might be launched as early as this month. This move comes as the firm accelerates efforts to leverage its growing prominence. The unveiling of DeepSeek-R1 sent shockwaves through the international technology sector due to its budget-friendly yet competitive performance compared to top-tier models.
DeepSeek has stayed mum regarding the rumored R2 release. They have refrained from commenting publicly on this issue via their official platforms, yet a customer support account reportedly dismissed the claim in a private conversation with corporate clients, according to reports by Chinese media outlets earlier last month.
DeepSeek did not promptly reply to requests for comments on Friday.
Although Hangzhou-based DeepSeek, established in 2023 by entrepreneur Liang Wenfeng For the last few months, which have seen the company thrust into the global limelight, it has mostly stayed out of the public eye, choosing instead to channel its efforts into research and development.
Last month, the firm updated its V3 model, known as DeepSeek-V3-0324, which it said offered "enhanced reasoning capabilities, optimised front-end web development and upgraded Chinese writing proficiency".
In February, it also made five of its code repositories open-source ,enabling developers to examine and add to its software development efforts. The startup committed to making "significant strides with complete openness."
In the same period, Liang released a detailed research paper focusing on "intrinsic sparse attention," which aims at enhancing the efficiency of LLMs when handling extensive datasets.
Liang, who is 40 years old, founded High-Flyer Quant, the parent company of DeepSeek. The hedge fund behind this enterprise has provided substantial financial support for the startup’s technological advancements.
In late February, the business owner participated in an event or activity. symposium with tech entrepreneurs Hosted by Chinese President Xi Jinping in Beijing, the launch was celebrated as an indication of China's perseverance despite US attempts to curb the nation's advancements in artificial intelligence.
More Articles from SCMP
Trump’s proposal for tariffs on vessels from China isn’t going to rescue U.S. shipbuilding facilities.
'Tai Chi in Fog': Why China's Objectives for International Arbitration Remain Elusive
2 people were hurt when a truck slid down an incline and collided with a vehicle in Hong Kong.
Hong Kong will host the Ultimate Tennis Showdown, featuring confirmed stars including Zhang and Rublev.
The article initially appeared on the South China Morning Post (www.scmp.com), which serves as the premier source for news coverage of China and Asia.
Copyright © 2025. South ChinaMorning Post Publishers Ltd. All rights reserved.
0 Komentar