A research team led by Alibaba Cloud was the only group from China to receive a top award at this year’s NeurIPS, or the Conference on Neural Information Processing Systems, billed as the artificial intelligence industry’s most prestigious annual event.
The research could lead to drastic improvements in the efficiency of large language models (LLMs), significantly reducing both training and inference costs for Alibaba Group Holding’s next generation of Qwen models without sacrificing accuracy.
The technique was rigorously validated through more than 30 experiments across models of varying sizes and architectures, demonstrating its robustness and generalisability.
Advertisement
The Alibaba Cloud team was one of four to be awarded best paper on Thursday – selected from 21,575 submissions – ahead of the conference opening on Sunday, with judges praising the tech giant for publishing its findings at a time when leading US players were increasingly keeping their AI research behind closed doors.
Alibaba Cloud is the AI and cloud computing unit of Alibaba, owner of the South China Morning Post.
Alibaba’s paper – which was co-authored by researchers from the University of Edinburgh, Stanford University, the Massachusetts Institute of Technology and Tsinghua University – introduced a new technique for improving AI foundational models’ “attention” mechanism, according to co-author and Alibaba Cloud chief technology officer Zhou Jingren.