Diffusion language models improve following text constraints in place
In-Place Instruction Following in Diffusion Language Models
Computation and LanguageArtificial Intelligence
Summary
Some models that generate text work by gradually refining noisy guesses and can handle specific instructions to edit text exactly where asked. The authors built a benchmark to see how well these models follow such precise instructions, like changing writing style or meaning in a specific spot. They found that standard models don’t focus enough on the parts where instructions apply. To fix this, they created a training method that makes the models better at obeying these in-place instructions without losing their overall language skills.
What this means in practice
- •For natural language processing engineers: Improve text generation systems to better follow exact user edits or instructions anchored at specific parts of the text during generation.
- •For content creation platforms: Enhance tools that assist writers by allowing precise style or discourse changes at targeted points in generated text.$Commercial implications: This enables content platforms to offer refined AI-assisted editing features that adjust style or meaning exactly where users want, improving user control.
Authors
Zheng Nie, Zherui Li, Jiaming Zhang, Kun Wang, Zhenhong Zhou, Yufei Guo
Abstract
Diffusion Large Language Models (dLLMs) generate text via bidirectional iterative denoising, naturally supporting user-specified constraints anchored at arbitrary output positions, a paradigm known as In-place Prompting (IPP). We formalize this as the In-place Instruction Following (IIF) task and construct IIF-Bench, a hierarchical benchmark spanning literal, style, and discourse-function constraints, paired with a rubric-based local-global evaluation protocol. An inference-time attention-bias probe suggests that vanilla dLLMs often under-prioritize constraint spans during denoising. We then propose GRAFT, an IPP-oriented post-training framework combining constraint-aware SFT and preference optimization. On four representative dLLMs, GRAFT raises the average IIF score from 57.75 to 73.10 (+15.35 points), with absolute gains of 15.91 and 15.57 points on literal and discourse-function constraints, while preserving general generation ability.