Peking University
academicPeking University's foundation-model output runs on two engines. Yaodong Yang's PKU-Alignment (Institute for AI) is the most complete academic safety/alignment stack anywhere: the Safe RLHF → Aligner (NeurIPS 2024 Oral) → Align-Anything line, the PKU-SafeRLHF preference data much of the field trains cost models on, and the ACL 2025 Best Paper showing language models elastically resist alignment.
Li Yuan's PKU-YuanGroup (Shenzhen Graduate School; HF org LanguageBind) ships the open video/multimodal line: Open-Sora Plan (12.2K★, the leading open Sora reproduction, with v1.5's 8B DiT pretrained entirely on Huawei Ascend), the foundational LanguageBind/Video-LLaVA lineage, and UniWorld unified understanding+generation. Supporting benches: Tong Yang's PKU-DS-LAB (FairyR1-32B branch-merge distillation, iFairy 2-bit complex-valued LLMs), Muhan Zhang's group (PiSSA, shipped in HF PEFT, and TransMLA), and Yisen Wang's PKU-ML (LongPPL).
Carve-outs: BIGAI-branded work (the CraftJarvis/MineStudio Minecraft-agent lineage under Yitao Liang) is excluded from this entry, as is BeingBeyond-branded humanoid work (Being-0; Being-H0 is joint). Joint-lab channels run through PKU–PsiBot (Yaodong Yang), PKU–Rabbitpre (Open-Sora Plan), and Qihoo360 (TinyR1 method paper).
People
- Yaodong Yang Google Scholar — Boya Assistant Professor & Assistant Dean, Institute for AI; PKU-Alignment lead; Chief Scientist, PKU–PsiBot joint lab
- Li Yuan Google Scholar — Assistant Professor, Shenzhen Graduate School; PKU-YuanGroup / Open-Sora Plan lead; PKU–Rabbitpre joint lab
- Muhan Zhang Google Scholar — Assistant Professor, Institute for AI; PiSSA & TransMLA (formerly Facebook AI Research)
- Tong Yang Google Scholar — Associate Professor, CS; PKU-DS-LAB (FairyR1, iFairy; TinyR1 joint with Qihoo360)
- Liwei Wang Google Scholar — Professor, ML theory; frequent Huawei Noah's Ark co-authorships