Foundation models & post-training
Research on long-context data selection, reasoning-path supervision, and model merging to improve reasoning efficiency and generalization.
Research on long-context data selection, reasoning-path supervision, and model merging to improve reasoning efficiency and generalization.
Research on dynamic attacks, safety alignment, and hierarchical policy control, balancing risk protection, task utility, and generation reliability.
Research on evidence loss in complex documents, multi-document evidence allocation, long-term memory, and experience transfer in realistic agent workflows.
Open 32B models spanning branch-merge distillation, safety control, and distribution-aware model merging for reasoning, instruction following, and safety.
Research and engineering across dynamic attack–defense, closed-loop training, and detection–response separation; the related project received the Beijing Science and Technology Progress Award, Second Prize.
Document understanding, retrieval, and trustworthy generation, from industrial OCR robustness to budgeted multi-document evidence allocation.
Advances Mem360 long-term memory and experience reuse, with research on agent workflows, memory safety, and experience serving.
An intelligent assistant for real-estate agents, combining search, question answering, and data intelligence; included in Beike’s U.S. IPO prospectus.
AI+VR work at Rushi/Beike, bringing speech and intelligent interaction to property narration and viewing.