The New Stage of Sino-US AI Competition, Tongyi Qianwen Wins the Best Paper at Top AI Conference
At this stage when Sino-US technological competition has fully entered a new phase, Chinese AI once again steps onto the global stage. This time, it's Tongyi Qianwen that has won the Best Paper Award at the top AI conference NeurIPS 2025.
In this year's NeurIPS, the Tongyi Qianwen team stood out from more than 20,000 submissions worldwide and became the only Chinese team to win the Best Paper Award this year. As one of the most influential AI conferences globally, NeurIPS has always been a barometer for AI technology development. Key technologies such as Transformer and AlexNet were born here. The acceptance rate for this year's submissions was only 25%, while there were only four best papers globally, with an acceptance probability of less than two ten-thousandths. On this stage where top research forces such as OpenAI, Google, and Meta are competing, the fact that a Chinese team has taken center stage itself indicates: Chinese teams have moved from participants to leaders.
More importantly, this award is not due to showing off skills but rather because it reveals a fundamental puzzle in the large model industry and publicly demonstrates an excellent solution. This paper is the first to systematically reveal how attention gate mechanisms affect the performance and training stability of large models, explaining a key question: Why can large models become more stable, more efficient, and smarter?
This breakthrough is regarded by the industry as an important progress in breaking through the current bottlenecks in large model training, which is expected to improve training efficiency and the performance of large models and drive the overall evolution of future large model architectures.
More importantly, this technology has already been applied to models such as Qwen3-Next, achieving a closed loop from academic innovation to engineering implementation, demonstrating the maturity of China's systematic innovation capabilities in AI. NeurIPS officials pointed out that this technology will be widely applied and promote AI researchers' understanding of attention mechanisms in large language models.
Currently, Qwen has open-sourced over 300 models, with global downloads exceeding 700 million times, and derivative models exceeding 180,000. It can be said that China's AI is not just publishing a top conference paper, but also influencing the world through open-source, allowing more developers to benefit.
This path of open innovation has enhanced China's voice in the global open-source community.
From the perspective of Sino-US competition, today's Sino-US AI competition is shifting from comparing model size and application landing to comparing underlying architecture and original innovation.
Whoever can achieve breakthroughs in architecture, mechanisms, and training principles will seize the initiative in the next stage.
The awarding of Tongyi Qianwen at NeurIPS demonstrates the true capabilities of Chinese teams in core areas: not following, not replicating, but achieving breakthroughs in fundamental innovation.
This is not only an achievement of Alibaba, but also a highlight moment for China's AI in global competition.
It symbolizes that China's AI is moving from engineering-driven catch-up to systematic originality; from a technical participant to a co-shaper of technical rules.
The Sino-US AI competition is entering a deeper and more challenging phase. And Chinese research teams are continuously securing key nodes one after another.
Original source: www.toutiao.com/article/1850126460022791/
Disclaimer: This article represents the views of the author.