Research map / Image generation / Fast diffusion sampling
DAPD: Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs
BumJun Kim, Dongjae Jeon, Moongyu Jeon, Albert No
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 |
|---|---|---|---|
| select independent set of attention-induced dependency graph for parallel unmasking instead of confidence/entropy/threshold selection | LLaDA-8B-Instruct · TriviaQA 5-question prompts | decoding steps: 124.4 (Fast-dLLM) -> 66.2 steps at similar accuracy (52.12 vs 52.08) | Improved |
| use DAPD instead of token-by-token confidence decoding | LLaDA-8B-Instruct · TriviaQA 5-question prompts | decoding steps: 256.0 -> 66.2 (3.87x) | Improved |
| use DAPD in single-block regime versus block-wise baselines | LLaDA-8B-Instruct · MBPP, IFEval | accuracy | Improved |
| use DAPD versus baselines on other tasks | LLaDA-8B-Instruct · HumanEval, GSM8K, Math500 | accuracy vs steps | Improved |
| apply DAPD to Dream-7B in single-block regime | Dream-7B-Instruct · code, math, IFEval | accuracy-steps trade-off | Improved |
| apply DAPD on ParallelBench | LLaDA-8B-Instruct · ParallelBench | score vs steps | Improved |
| run baselines in single-block decoding without block-wise decoding or EOS suppression (ablation) | Fast-dLLM, KLASS, EB-Sampler on LLaDA · code, math, IFEval | accuracy | Hurt |
| use attention edge scores from top layers on synthetic MRF (ablation) | attention graph construction · Synthetic MRF toy data | AUC: AUC 0.928, ratio 2.204, OVR 0.04 | Improved |
| switch to confidence-threshold (>0.9) rule when mask ratio falls below 50% (ablation) | DAPD · decoding | decoding steps | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 2 on most shared results.
- Fast-dLLM — Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding (ICLR 2026) · better on 1 of 2 shared results
- KLASS — KLASS: KL-Guided Fast Inference in Masked Diffusion Models (NeurIPS 2025) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DMax: Aggressive Parallel Decoding for dLLMs (NeurIPS 2026 · similarity 0.91)
- Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.89)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.89)
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.
