Xin Cheng

Worth tracking if you care about alternatives to the standard transformer playbook, especially the line of work trying to keep strong language-model performance while making inference and memory use much cheaper.

A distinctive page because his work bridges open-sequence-model experimentation with applied machine learning for molecules, proteins, and structural biology, and he shows up on multiple RWKV-family papers including the hybrid GoldFinch branch rather than only the first release.

A strong open-model and data-centric page because his work sits close to the infrastructure that made OLMo and Dolma useful to the broader research community rather than just another benchmark-driven model release.

Co-authored RWKV: Reinventing RNNs for the Transformer Era.

Useful because it turns an otherwise thin RWKV byline into a real systems profile: after the original paper, his public work tracks toward large-scale pretraining infrastructure, pipeline parallelism, and systems support for frontier-scale models.

A useful page because it turns an otherwise stray RWKV byline into a visible builder profile: his public work is less about academic publishing and more about making efficient models, AI agents, and production RWKV systems usable in practice.