AIPrompt-infilling paper for diffusion language models accepted to COLM 2026 workshop
PatronusAI says masking both prompts and responses during supervised fine-tuning enables diffusion language models to infer task-adapted prompts from a few examples.
TLDR
PatronusAI announced that its paper was accepted to COLM 2026’s “Non-Autoregressive Language Models for Fast & Flexible Text Generation” workshop. The team says models such as LLaDA and Dream cannot infer prompts from desired outputs out of the box, and attributes that limitation to prompts never being masked during supervised fine-tuning. It says masking both prompts and responses—called full-sequence masking—unlocks that ability. PatronusAI is calling for the community to release models fine-tuned this way alongside standard versions.
