Research map / Image generation / Fast diffusion sampling
DARL: Efficient Document-to-Markup Generation via Look-Ahead Diffusion Trajectory Sampling
Wentao Yang, Yongxin Shi, Rui Tang, Peirong Zhang, Shihang Wu, Huiguo He, Zheng Huang, Dezhi Peng, Minghui Liao, Lianwen Jin
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 Sliding Diffusion Block (ablation) | dots.ocr-3B · OmniDocBench-1.5 | wall-time speedup: 1.00× -> 1.98× | Improved |
| add Sliding Diffusion Block (ablation) | dots.ocr-3B · OmniDocBench-1.5 | Overall score: 87.57 -> 87.21 | Hurt |
| add immediate reward r1 (ablation) | DARL (SDB only) · OmniDocBench-1.5 | Overall score: 87.21 -> 87.53 | Improved |
| add immediate reward r1 (ablation) | DARL (SDB only) · OmniDocBench-1.5 | wall-time speedup: 1.98× -> 2.14× | Improved |
| add look-ahead reward r2 (ablation) | DARL (SDB + r1) · OmniDocBench-1.5 | Overall score: 87.53 -> 87.69 | Improved |
| add look-ahead reward r2 (ablation) | DARL (SDB + r1) · OmniDocBench-1.5 | wall-time speedup: 2.14× -> 2.31× | Improved |
| use LCP reward instead of edit distance (ablation) | DARL · OmniDocBench-1.5 | wall-time speedup: 2.13× -> 2.31× | Improved |
| use LCP reward instead of edit distance (ablation) | DARL · OmniDocBench-1.5 | Overall score: 87.85 vs 87.69 | No clear effect |
| use block size 16 instead of larger blocks (ablation) | DARL · OmniDocBench-1.5 | Overall score: 87.69 -> 82.91 | Improved |
| add DARL to Qwen3-VL-2B | Qwen3-VL-2B · OmniDocBench-1.5 | Overall score: 82.47 -> 82.24 | No clear effect |
| add DARL to Qwen3-VL-2B | Qwen3-VL-2B · OmniDocBench-1.5 | wall-time speedup: 1.00× -> 2.33× | Improved |
| scale InternVL3.5 with DARL from 2B to 14B | InternVL3.5 +DARL · OmniDocBench-1.5 | wall-time speedup: 2.20× -> 2.78× | Improved |
Closest papers by meaning
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.87)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.86)
- Energy-Based Diffusion Language Models for Text Generation (ICLR 2025 · similarity 0.85)
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.85)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.85)
- Dynamically Structured Diffusion Language Model Decoding via Bayesian Inference (NeurIPS 2026 · similarity 0.85)
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.
