Tencent's new AI algorithm gets accepted by top machine learning conference

Tencent's Hunyuan team introduced a new AI algorithm called Stem, which has been accepted by the top machine learning conference ICML-26. The algorithm uses only 25% of the computing power to achieve accuracy close to dense attention, thanks to two key innovations: Token Position Decay (TPD) and Output-Aware Metric (OAM).

With the help of an open-source HPC operator library, Stem delivers real performance gains. It reduces the first-token latency by 3.6 times for 128K context length. The paper and code are now available on GitHub, offering a new way to make large AI models run more efficiently.