Research map / Image generation / Fast diffusion sampling
Edit-Based Refinement for Parallel Masked Diffusion Language Models
Houxing Ren, Mingjie Zhan, Zimu Lu, Ke Wang, Yunqiao Yang, Haotian Hou, Junting Pan, Hongsheng Li
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 edit-based refinement stage (Stage-3) on top of Stage-2 | ME-DLM · GSM8K (budget 1/8) | accuracy: 30.2 -> 63.8 (+33.6) | Improved |
| add edit-based refinement stage (Stage-3) on top of Stage-2 | ME-DLM · HumanEval (budget 1/8) | pass@1: 13.4 -> 25.0 (+11.6) | Improved |
| add edit-based refinement stage (Stage-3) on top of Stage-2 | ME-DLM · Average (budget 1/1) | average: 55.7 -> 60.0 (+4.3) | Improved |
| add edit-based refinement stage (Stage-3) on top of Stage-2 | ME-DLM · Average (budget 1/8) | average: 19.3 -> 32.6 (+13.3) | Improved |
| SFT with diffusion-style unmasking (Stage-2) vs LLaDA-Instruct | ME-DLM · Average (budget 1/8) | average: 20.9 -> 19.3 | Mixed |
| add model-generated refinement states (vs mask-only states) (ablation) | ME-DLM Stage-3 · Average (budget 1/4) | average: 41.9 -> 46.4 | Improved |
| add model-generated refinement states (vs rule-based noise) (ablation) | ME-DLM Stage-3 · Average (budget 1/8) | average: 21.4 -> 32.6 | Improved |
| add model-generated refinement states (vs mask-only states) (ablation) | ME-DLM Stage-3 · Average (budget 1/1) | average: 57.4 -> 60.0 | Improved |
| restrict edit noise range (beta=0.5 vs 0) (ablation) | ME-DLM Stage-3 · Average (budget 1/8) | average: 32.6 -> 39.1 | Improved |
| restrict edit noise range (beta=0.5 vs 0) (ablation) | ME-DLM Stage-3 · Average (budget 1/1) | average: 60.0 -> 57.0 | Hurt |
| vary proportion alpha of mask-edit training (0.25, 0.75 vs 0.5) (ablation) | ME-DLM Stage-3 · Average (budget 1/1) | average: 60.0 vs 59.2 vs 59.6 | No clear effect |
| add edit diffusion steps at inference (vs mask-only) (ablation) | ME-DLM · MATH-500 (512 steps) | accuracy: 45.8 -> 49.6 | Improved |
Closest papers by meaning
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- Beyond Masks: Efficient, Flexible Diffusion Language Models via Deletion-Insertion Processes (ICLR 2026 · similarity 0.92)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.90)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
- Theoretical Benefit and Limitation of Diffusion Language Model (NeurIPS 2025 · similarity 0.90)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.90)
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.
