Research map / Large language models / Language, safety and interpretability
Dynamic-dLLM: Dynamic Cache-Budget and Adaptive Parallel Decoding for Training-Free Acceleration of Diffusion LLM
Tianyi Wu, Xiaoxi Sun, Yanhua Jiao, Yulin Li, Yixin Chen, Yun-Hao Cao, Yi-Qi Hu, Zhuotao Tian
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 Dynamic-dLLM with parallel decoding | LLaDA-8B-Instruct · GSM8K | throughput (TPS): 8.32 -> 37.29 | Improved |
| apply Dynamic-dLLM with parallel decoding | LLaDA-8B-Instruct · GSM8K | accuracy: 78.21 -> 78.01 | No clear effect |
| apply Dynamic-dLLM instead of Fast-dLLM, with parallel decoding | LLaDA-8B-Instruct · Average of 5 benchmarks | throughput speedup: 2.85x -> 3.21x | Improved |
| apply Dynamic-dLLM with parallel decoding | Dream-v0-7B-Instruct · Average of 5 benchmarks | accuracy: 66.27 -> 65.01 | Hurt |
| increase cache budgets B_window and B_layer from 32 to 128 (ablation) | Dynamic-dLLM · GSM8K (1024 tokens) | accuracy: 73.92 -> 79.15 | Improved |
| remove layer budget (B_layer=0) (ablation) | Dynamic-dLLM · GSM8K (1024 tokens) | accuracy: 74.07 vs 77.62 at 64/64 budget | Hurt |
| scale alpha and beta by 100x (ablation) | Dynamic-dLLM · GSM8K (256 tokens) | accuracy: 78.01 -> 59.76 | Hurt |
| use dynamic instead of fixed confidence threshold (ablation) | Dynamic-dLLM · GSM8K | inference steps: approx. 30% fewer steps | Improved |
| apply Dynamic-dLLM on RTX 4090 | LLaDA-8B-Instruct · GSM8K | throughput (TPS): 5.48 -> 29.77 | Improved |
Closest papers by meaning
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.93)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.92)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
- Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.92)
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.
