Research map / Image generation / Fast diffusion sampling
dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching
Zhiyuan Liu, Yicun Yang, Yaojie Zhang, Junjie Chen, Chang Zou, Qingyan Wei, Shaobo Wang, Yichen Zhu, Linfeng Zhang
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 |
|---|---|---|---|
| add dLLM-Cache | LLaDA Base · GSM8K | score: 69.06 -> 70.66 with 5.02x FLOPs speedup | Improved |
| add dLLM-Cache | Dream Instruct · HumanEval | score: 57.92 -> 56.09 | Hurt |
| add dLLM-Cache | LLaDA Instruct · GPQA | FLOPs: 22.07T -> 2.73T, 8.08x | Improved |
| add dLLM-Cache | LLaDA and Dream Instruct · LongBench | average score: 39.14 -> 38.61 (LLaDA); 33.04 -> 32.70 (Dream) | Hurt |
| use dLLM-Cache instead of dKV-Cache or Fast-dLLM | Dream Base · GPQA | score: 31.31 -> 34.15 vs Fast-dLLM; 32.83 vs dKV-Cache | Improved |
| combine dLLM-Cache with SlowFast Sampling | LLaDA Base · GSM8K | TPS: 7.37 -> 49.86 TPS, 67.17 accuracy | Improved |
| reduce denoising steps to 32 instead of caching | LLaDA Base · GSM8K | accuracy: 69.06 -> 22.25 | Hurt |
| overwrite full V cache instead of selected tokens only (ablation) | V-verify · GSM8K | accuracy: 78.85 -> 79.83 | Improved |
| use cosine similarity instead of L2 distance (ablation) | V-verify · GSM8K | accuracy: 55.95 -> 78.54 | Improved |
| use fixed global rho 0.25 instead of lower-layer-heavy allocation (ablation) | dLLM-Cache · GSM8K | accuracy: 78.70 -> 79.83 | Improved |
| use fixed global rho 0.25 instead of higher-layer-heavy allocation (ablation) | dLLM-Cache · GSM8K | accuracy: 79.23 -> 79.83 | Improved |
| increase response refresh interval Kr from 2 to 6 (ablation) | dLLM-Cache · GPQA (Dream 7B) | accuracy: 36.16 -> 33.92 | Hurt |
Later papers that compared with this method
2 later accepted papers report + dLLM-Cache as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- d$^2$Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching (ICLR 2026 · better on 7 of 7 shared results)
- Attention Is All You Need for KV Cache in Diffusion LLMs (ICLR 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 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 2 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
- Dynamic-dLLM: Dynamic Cache-Budget and Adaptive Parallel Decoding for Training-Free Acceleration of Diffusion LLM (ICLR 2026 · similarity 0.94)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.94)
- d$^2$Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching (ICLR 2026 · similarity 0.94)
- EntropyCache: Decoded Token Entropy Guided KV Caching for Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models (ICML 2025 · 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.
