Research map / Large language models / Language, safety and interpretability
dParallel: Learnable Parallel Decoding for dLLMs
Zigeng Chen, Gongfan Fang, Xinyin Ma, Ruonan Yu, Xinchao Wang
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 |
|---|---|---|---|
| apply certainty-forcing distillation (dParallel) | LLaDA-8B-Instruct · GSM8K-CoT (0-shot) | steps / accuracy: 256 -> 30 steps; 75.7% -> 76.1% | Improved |
| apply certainty-forcing distillation (dParallel) | LLaDA-8B-Instruct · MBPP (3-shot) | steps / accuracy: 256 -> 24 steps; 42.4% -> 40.8% | Improved |
| apply certainty-forcing distillation (dParallel) | Dream-7B-Instruct · GSM8K-CoT (0-shot) | speedup: 1.0x -> 6.9x; 82.9% -> 82.1% | Improved |
| cut decoding steps to 64 without training (few-step decoding) | LLaDA-8B-Instruct · HumanEval (0-shot) | accuracy: 38.4% -> 19.5% | Hurt |
| use consistency distillation to cut steps to 64 | LLaDA-8B-Instruct · MBPP (3-shot) | accuracy: 42.4% -> 25.0% | Hurt |
| add dual KV-cache | LLaDA-8B-Instruct · HumanEval (0-shot) | accuracy: 38.4% -> 34.1% | Hurt |
| add certainty loss on top of consistency loss (ablation) | dParallel · GSM8K-CoT (0-shot) | steps: 53 -> 30 steps; 73.5% -> 76.1% | Improved |
| add consistency loss on top of certainty loss (ablation) | dParallel · GSM8K-CoT (0-shot) | accuracy: 57.8% -> 76.1% | Improved |
| add semi-autoregressive forward masking (ablation) | dParallel · GSM8K-CoT (0-shot) | steps: 44 -> 30 steps; 73.3% -> 76.1% | Improved |
| use fixed 50% masking ratio instead of random or other ratios (ablation) | dParallel · GSM8K (LLaDA, 1 epoch) | accuracy: 72.4% -> 76.3% (random) | Improved |
| use semi-AR masking when training Dream (ablation) | Dream-7B-Instruct · training | degeneration risk | Hurt |
| use dParallel instead of confidence-threshold decoding at equal speedup | LLaDA-8B-Instruct · GSM8K at 9.4x speedup | accuracy: +16.5% | Improved |
Closest papers by meaning
- DMax: Aggressive Parallel Decoding for dLLMs (NeurIPS 2026 · similarity 0.92)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.91)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.91)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.90)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
- PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion 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.
