Research
We partner with companies and research teams on high-difficulty problems in agentic AI — world models, embodied reasoning, and autonomous systems — and deliver publishable, peer-reviewed research.
From Partially to Fully Unmatched Modalities as Negative Samples in Contrastive Learning
We introduce a curriculum for multimodal contrastive learning that treats partially-to-fully unmatched modality pairs as progressively harder negative samples, improving cross-modal representation quality.
Toward an Open Ecosystem for Federated Learning Mobile Apps
FLSys is a mobile-cloud federated learning system that supports different FL models and aggregation methods, enabling an open ecosystem of privacy-preserving FL apps on smartphones.
Low-Overhead Model Pruning for Federated Learning
Complement Sparsification prunes models through complementary, collaborative sparsification at the server and the clients, cutting communication and computation overhead in federated learning.
