Kaidi Cao is a member of technical staff at Anthropic, working on Claude’s agentic and tool-use capabilities. He previously co-founded Voyage AI and worked on embedding and reranking models for search.[1]
Start with Label-Distribution-Aware Margin Loss →
GitHub Career
Cao holds a BE from Tsinghua and a PhD in computer science from Stanford. His research spans learning with imbalanced data, graph models and representations for scientific data. At Voyage AI, he helped build embedding and reranking models before the company’s acquisition. His later focus at Anthropic concerns how Claude uses tools and acts through an agent environment.[1][2][3][4]
Contributions
Learning from less frequent examples
Cao is first-listed co-author of Label-Distribution-Aware Margin Loss. The work addresses training sets in which some classes have many more examples than others. It changes the classification margin according to class frequency and separately studies deferring reweighting until a representation has begun to form. These choices target generalization on less frequent classes without relying only on repeated sampling of rare examples.[2]
Keeping spatial context in cell annotation
He co-authored STELLAR, a geometric deep-learning method for annotating spatially resolved single-cell data. The method uses molecular and spatial information to transfer labels from a reference dataset, identify cell types already represented there and discover new types or states. Preserving tissue organization supplies information that would be lost if every cell were treated as an isolated measurement.[3]
Training on a graph in segments
Graph Segment Training divides a large graph into smaller pieces and backpropagates through a sampled subset each iteration. A table of historical embeddings supplies information for the remaining segments. The method also addresses stale representations through prediction-head fine-tuning and embedding dropout. It aims to make graph-level prediction possible within bounded training memory.[4]
Tools, writing & talks
Writing, research & talks
Start here
Label-Distribution-Aware Margin Lossarxiv.orgA classification loss and training schedule designed to improve learning when class frequencies are highly uneven.
Read the introduction’s class-imbalance problem before the margin derivation. Familiarity with supervised classification and cross-entropy helps; the repository linked from the paper provides an implementation.
First-listed co-author[2]- Graph Segment Trainingarxiv.org
A divide-and-conquer approach for training large-graph predictors within a bounded memory budget.
Co-author[4]