Zhilin Yang

Yang leads Moonshot AI. He completed a PhD at Carnegie Mellon in 2019, advised by Ruslan Salakhutdinov and William W. Cohen, after receiving a bachelor’s degree from Tsinghua in 2015. His academic biography also records work at Meta AI and Google Brain. At NVIDIA’s 2026 GTC, he presented Moonshot’s account of scaling Kimi K2.5 through training-system design and model architecture.[1][2][6][7]

Learn bidirectional context through autoregressive training

XLNet trains over permutations of a sequence’s factorization order, allowing tokens to learn from different contexts without the masked-token corruption used by BERT. Yang and Zihang Dai are marked equal contributors. The paper combines this objective with ideas from Transformer-XL and evaluates the resulting pretrained model across language-understanding tasks.[3]

Carry information across text segments

Transformer-XL introduces segment-level recurrence and relative positional information so a language model can use information beyond a fixed input segment. Its design addresses both limited context and fragmentation at segment boundaries. Dai and Yang are marked equal contributors; the work is a collaboration with Yiming Yang, Jaime Carbonell, Quoc V. Le and Ruslan Salakhutdinov.[4]

Scale an existing optimizer

Yang coauthored the Moonlight report on scaling Muon, an optimizer that predates that work. The team added weight decay and adjusted update scale, then evaluated the approach in large-language-model training. The report releases a distributed implementation and model checkpoints, making both the optimization method and resulting training runs available for further study.[5]