Research map / Large language models / Language, safety and interpretability
ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMs
Wonjun Kang, Kevin Galim, Seunghyuk Oh, Minjae Lee, Yuchen Zeng, Shuibai Zhang, Coleman Richard Charles Hooper, Yuezhou Hu, Hyung Il Koo, Nam Ik Cho, Kangwook Lee
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 |
|---|---|---|---|
| fine-tune LLaDA 1.5 on Waiting Line, then decode Replace Index in parallel | LLaDA 1.5 · Waiting Line Replace Index | accuracy: below 50% -> nearly 100% | Improved |
| fine-tune LLaDA 1.5 and decode Replace Random and Shuffle in parallel | LLaDA 1.5 · Waiting Line Replace Random / Shuffle | accuracy under parallel decoding | No clear effect |
| add Chain-of-Thought prompting | LLaDA 1.5 · Waiting Line | accuracy under parallelism: 8x more output tokens | Mixed |
| use Confidence Threshold instead of Top-k at conservative thresholds | LLaDA 1.5 · ParallelBench | speed-quality trade-off | Improved |
| use Confidence Threshold with aggressive thresholds | LLaDA 1.5 · ParallelBench | accuracy | Hurt |
| use Confidence Top-k instead of Random Top-k on tasks with C(Y|X)=0 (Reverse, Replace Index) | LLaDA 1.5 · Waiting Line | accuracy | Improved |
| use Confidence Top-k instead of Random Top-k on tasks with C(Y|X)>0 (Replace Random, Shuffle) | LLaDA 1.5 · Waiting Line | accuracy | Hurt |
| use small semi-AR blocks (more left-to-right) | LLaDA 1.5 (Random Top-2) · Text Writing vs Waiting Line | accuracy | Mixed |
| use oracle per-sample threshold instead of fixed threshold | Confidence Threshold · ParallelBench | accuracy and speedup | Improved |
| increase parallelism k from 1 to larger on Shuffle | LLaDA 1.5 · Shuffle | accuracy: near-perfect -> zero | Hurt |
Closest papers by meaning
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.93)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.92)
- DMax: Aggressive Parallel Decoding for dLLMs (NeurIPS 2026 · similarity 0.91)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.91)
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.
