Research map / Architectures and efficiency / Attention and transformers
Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction
Yuerong Song, Xiaoran Liu, Ruixiao Li, Zhigeng Liu, Zengfeng Huang, Qipeng Guo, Ziwei He, Xipeng Qiu
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 Sparse-dLLM cache eviction | LLaDA-8B-Instruct · GSM8K | throughput (TPS): 4.57 -> 26.45 | Improved |
| apply Sparse-dLLM cache eviction | Dream-v0-7B-Instruct · GPQA | throughput (TPS): 6.29 -> 32.63 | Improved |
| apply Sparse-dLLM cache eviction | LLaDA-8B-Instruct · 7-benchmark average | accuracy (%): 59.24 -> 59.86 | No clear effect |
| apply Sparse-dLLM cache eviction | Dream-v0-7B-Instruct · HumanEval | accuracy (%): 57.93 -> 50.00 | Hurt |
| apply Sparse-dLLM cache eviction | Dream-v0-7B-Instruct · peak memory | peak memory (GB): 0.9x of vanilla | Improved |
| apply Sparse-dLLM cache eviction | LLaDA-8B-Instruct · LongBench | accuracy (%): 34.55 -> 34.46 | No clear effect |
| add bidirectional (prefix and suffix) eviction instead of prefix-only (ablation) | Sparse-dLLM on LLaDA-8B-Instruct · GSM8K/MATH/ARC-C average | accuracy (%): 65.40 -> 66.82 | Improved |
| add bidirectional eviction instead of prefix-only (ablation) | Sparse-dLLM on Dream-v0-7B-Instruct · GSM8K/MATH/ARC-C average | accuracy (%): 69.04 -> 69.04 | No clear effect |
| increase cache update delay from 1 to 5 steps (ablation) | Sparse-dLLM · ARC-C | throughput (TPS): 36.85 -> 33.78 | Hurt |
| increase delay step from 1 to 3 (ablation) | Sparse-dLLM · ARC-C | accuracy (%): 88.47 -> 89.49 | Mixed |
| raise retention ratio above 0.5 (ablation) | Sparse-dLLM · GSM8K | accuracy (%) | No clear effect |
| change max-pool kernel size away from 3 (ablation) | Sparse-dLLM · GSM8K | accuracy (%) | Hurt |
Closest papers by meaning
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
- EntropyCache: Decoded Token Entropy Guided KV Caching for Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.89)
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.
