Releases, papers, SOTA benchmarks
A new RNN architecture outperforms Transformers by treating each hidden state as an independent model. The first author claims this approach fundamentally changes language modeling.
Kimi paper reveals inference architecture handling 80% of traffic: AI forward-looking topic (batch3, outside main timeline Jan 2025–May 2026).
Kaiming He leads his first team post-MIT hiring on a new AI generation project, with double Olympiad gold medalist Mingyang Deng participating. This work falls outside the main 2025-2026 publication timeline.
A new pure MLP architecture from Tsinghua University and Ant Group outperforms Transformers in short- and long-term time series forecasting.
A new domestic generative AI model matches AlphaFold3 in performance, simultaneously predicting antigen-antibody complex structures and enabling de novo antibody design.
BaiChuan unveils a new model that tops Chinese benchmarks alongside its first AI assistant, Bai Xiaoying, dubbed the most search-savvy. This falls outside the main 2025-2026 timeline.