Research map / Image generation / Fast diffusion sampling
DreamOn: Diffusion Language Models For Code Infilling Beyond Fixed-size Canvas
Zirui Wu, Lin Zheng, Zhihui Xie, Jiacheng Ye, Jiahui Gao, Shansan Gong, Yansong Feng, Zhenguo Li, Wei Bi, Guorui Zhou, Lingpeng Kong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DreamOn length-control states to fine-tune Dream-7B | Dream-7B · HumanEval-Infilling single-line | pass@1: 48.2 -> 88.6 | Improved |
| add DreamOn to DreamCoder-7B (mask length 64) | DreamCoder-7B · HumanEval-Infilling multi-line | pass@1: 43.2 -> 63.8 | Improved |
| add DreamOn to DreamCoder-7B | DreamCoder-7B · SantaCoder-FIM Python | exact match: 59.3 -> 79.0 | Improved |
| add [delete] state (contraction) to DreamOn (ablation) | DreamOn · HumanEval-Infilling single-line (avg over mask lengths) | pass@1: 67.4 -> 90.8 | Improved |
| add [expand] state to DreamOn (ablation) | DreamOn · HumanEval-Infilling single-line (avg over mask lengths) | pass@1: 73.4 -> 90.8 | Improved |
| add [expand] state to DreamOn (ablation) | DreamOn · HumanEval-Infilling multi-line (avg over mask lengths) | pass@1: 32.4 -> 57.1 | Improved |
| add [delete] state to DreamOn (ablation) | DreamOn · SantaCoder-FIM Python (avg over mask lengths) | exact match: 53.0 -> 77.4 | Improved |
| down-weight loss on [delete] tokens (loss balancing) (ablation) | DreamOn · HumanEval-Infilling single-line (avg over mask lengths) | pass@1: 84.6 -> 90.8 | Improved |
| remove [delete] tokens in place instead of keeping placeholders (ablation) | DreamOn · HumanEval-Infilling single-line (avg over mask lengths) | pass@1: 84.6 -> 90.8 | Improved |
| add inference-time deletion broadcasting (ablation) | DreamOn · HumanEval-Infilling single-line (avg over mask lengths) | pass@1: 90.2 -> 90.8 | Improved |
| add deletion broadcasting for efficiency (ablation) | DreamOn · generation steps at mask length 64 | inference steps: 122.8 -> 52.4 | Improved |
| use DreamOn vs SFT on ROCStories story ending at mask length 8 | Dream-7B · ROCStories story ending | Rouge-L: 25.1 -> 24.6 | No clear effect |
Closest papers by meaning
- Diffusion Language Models Can Approximate Optimal Infilling Lengths Implicitly (NeurIPS 2026 · similarity 0.91)
- Diffusion-State Policy Optimization for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.88)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.87)
- Theoretical Benefit and Limitation of Diffusion Language Model (NeurIPS 2025 · similarity 0.87)
- Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata (NeurIPS 2026 · similarity 0.87)
- LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
