ACL 2026 · Long Paper · Oral Presentation

Unlocking Diffusion Language Models through Template Infilling

Junhoo Lee · Seungyeon Kim · Nojun Kwak — Seoul National University

Stop forcing diffusion language models to imitate autoregressive prompting. Give them a global response blueprint, then let them fill it in.

+9.40ppaverage gain over corresponding baselines
4DLM model variants
0 GPUfor this pre-rendered oral demo

From the oral deck

Watch structure persist through denoising

This is the actual TI-only guardrail animation used in our ACL oral presentation. The response canvas reserves a draft, a safety critique, and a final safe response before generation finishes.

Template Infilling guardrail generation from the ACL oral deck

Pre-rendered research example; no prompt or user data is sent anywhere.

Core idea

From one prefix to a whole-response blueprint

1 · Place anchorsDistribute persistent structural checkpoints across the target response.
2 · Fill in parallelMasked spans see both earlier context and future anchors during denoising.
3 · Expand when neededDynamic Segment Allocation gives uncertain spans extra reasoning room.
Template Infilling and Dynamic Segment Allocation method overview

Evidence

One conditioning method, several task families

+9.40ppaverage improvement over the corresponding vanilla baselines across the four reported model variants.

The paper evaluates LLaDA-8B and Dream-7B base/instruct variants on GSM8K, MATH500, HumanEval, and Trip Planning. TI consistently improves the reported settings while retaining native DLM flexibility.

Main benchmark table for Template Infilling

Citation

Use Template Infilling in your work

@inproceedings{lee-etal-2026-unlocking,
  title     = {Unlocking the Potential of Diffusion Language Models through Template Infilling},
  author    = {Lee, Junhoo and Kim, Seungyeon and Kwak, Nojun},
  booktitle = {Proceedings of ACL 2026 (Volume 1: Long Papers)},
  year      = {2026},
  doi       = {10.18653/v1/2026.acl-long.284},
  url       = {https://aclanthology.org/2026.acl-long.284/}
}