Research map / Large language models / Language, safety and interpretability
Hierarchy Decoding: A Training-free Parallel Decoding Strategy for Diffusion Large Language Models
Xiaojing Qi, Lun Du, Xinyuan Zhang, Lanning Wei, Tao Jin, Da Zheng
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 |
|---|---|---|---|
| use Hierarchy-dLLM hierarchical decoding instead of Fast-dLLM | LLaDA-Instruct-8B · GSM8K | TPS: 12.80 -> 19.41 | Improved |
| use Hierarchy-dLLM instead of WINO | LLaDA-Instruct-8B · GSM8K | score (accuracy): 68.75 -> 77.18 | Improved |
| use Hierarchy-dLLM instead of Fast-dLLM | Dream-7B · Math500 | score (accuracy): 12.0 -> 16.0 | Improved |
| use Hierarchy-dLLM instead of vanilla decoding | Dream-7B · GSM8K | score (accuracy): 75.8 -> 73.2 | Hurt |
| use Hierarchy-dLLM instead of vanilla decoding | LLaDA-Instruct-8B · MBPP | TPS: 0.65 -> 11.20 | Improved |
| increase generation length from 256 to 1024 (ablation) | Hierarchy-dLLM · GSM8K | speedup over vanilla TPS: 4.49x -> 11.92x | Improved |
| increase block length from 16 to 64 (ablation) | Hierarchy-dLLM · GSM8K | TPS speedup: 5.81x -> 8.68x | Improved |
| increase block length from 32 to 64 (ablation) | Hierarchy-dLLM · GSM8K | score (accuracy): 83.70 -> 81.35 | Hurt |
| decrease block length from 32 to 16 (ablation) | Hierarchy-dLLM · GSM8K | score (accuracy): 83.70 -> 81.58 | Hurt |
| raise the high threshold tau_high (ablation) | Hierarchy-dLLM · GSM8K | TPS | Hurt |
| vary the low threshold tau_low (ablation) | Hierarchy-dLLM · GSM8K | TPS | No clear effect |
| use sparse masking instead of consecutive masking (ablation) | LLaDA (preliminary study) · HumanEval and GSM8K | approximate KL divergence | Improved |
Closest papers by meaning
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.92)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- Dynamically Structured Diffusion Language Model Decoding via Bayesian Inference (NeurIPS 2026 · similarity 0.90)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.90)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (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.
