Stable-DiffCoder is a code-focused diffusion language model that learns to write and edit programs by filling in masked pieces, not just predicting the next token.
The paper introduces UCoder, a way to teach a code-generating AI to get better without using any outside datasets, not even unlabeled code.