Plain guides to the research problems I work on. Each post explains one idea and walks through the key papers in that area (including some of mine).
Published:
A friendly tour of the tricks that make language models smarter after training. We cover thinking longer and voting and graders and self-checks. Then we look at a hidden step that can flip the result.
Published:
Modern biological machine learning increasingly relies on pretrained representations. A pathology foundation model can transform an image into rich embeddings, while a perturbation model can learn recurring transcriptional programs from large single-cell datasets. But representations alone do not always contain enough information to generalize reliably.