Self-Generated Error Training for Token Editing in Diffusion Language Models

In three linesTraining method to improve token editing in diffusion language models (LLaDA2.1). Addresses training-inference mismatch between random corruptions and model's own errors. Uses no-gradient draft pass followed by supervision on self-generated corruptions via LoRA. Reduces edit intensity and transcription errors.

Summary generated by Claude — human-verified