The paper, titled “Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention,” was published on February 27, with Liang listed as one of 15 authors. The “native sparse attention” mechanism is a core improvement that underpins the high efficiency and low-cost performance of DeepSeek’s AI models.
The paper’s win comes as Chinese scientists and researchers are outperforming US peers in basic research in the field of computational linguistics and natural language processing.
At this year’s ACL conference, more than half of the first-named authors on accepted papers originated from China, up from less than 30 per cent last year. The US ranked second, with 14 per cent of first-named authors, according to ACL data.
Among the four best papers recognised by ACL, two author teams were from China. They included Liang’s DeepSeek team and Yang Yaodong’s team from Peking University.
Yang, an assistant professor at the Institute of Artificial Intelligence and chief scientist of the Peking University-PsiBot Joint Laboratory, led research that explored a possible mechanism explaining the fragility of alignment in language models, attributed to the elasticity of language models.