Research map / Image generation / Fast diffusion sampling
Attention Is All You Need for KV Cache in Diffusion LLMs
Quan Nguyen-Tri, Mukul Ranjan, Zhiqiang Shen
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 |
|---|---|---|---|
| Elastic-Cache adaptive KV refresh instead of Fast-dLLM block caching | LLaDA-Instruct · GSM8K 5-shot 512 | accuracy (%) / throughput: 74.83 -> 77.71 acc; 44.0 -> 90.1 t/s | Improved |
| Elastic-Cache instead of fixed-interval DeepCache | LLaDA-1.5 · GSM8K 5-shot 512 | accuracy / throughput: 81.4 -> 83.7 acc; 60.9 -> 139.4 t/s | Improved |
| Raise attention threshold gamma from 0.5 to 0.95 (ablation) | Elastic-Cache · GSM8K 512, LLaDA-1.5 | accuracy (%) vs speedup: 76.04 -> 83.02 acc; 54.9x -> 37.8x speedup | Mixed |
| Raise gamma from 0.9 to 0.95 (ablation) | Elastic-Cache · GSM8K 512, LLaDA-Instruct | accuracy (%): 77.71 -> 76.72 | Hurt |
| Increase generation length from 256 to 1024 (ablation) | Elastic-Cache · GSM8K 5-shot, LLaDA | speedup over Fast-dLLM: 1.1x -> 6.8x | Improved |
| Increase few-shot prefill from 3 to 8 (ablation) | Elastic-Cache · GSM8K 1024, LLaDA | throughput (t/s): 185.3 -> 143.9 | Hurt |
| Use sliding window instead of block-wise decoding (ablation) | Elastic-Cache · GSM8K, LLaDA | accuracy | Improved |
| Increase sliding window size beta beyond about 64 (ablation) | Elastic-Cache · GSM8K, LLaDA | accuracy | Hurt |
| More aggressive parallel decoding (4 tok/step) (ablation) | Fast-dLLM · GSM8K, LLaDA-1.5 | accuracy (%): 80.5 -> 64.7 (1 to 4 tok/step) | Hurt |
| More aggressive parallel decoding (4 tok/step) (ablation) | Elastic-Cache · GSM8K, LLaDA-1.5 | accuracy (%): 82.6 -> 69.9 (1 to 4 tok/step) | Hurt |
| Elastic-Cache on code tasks with LLaDA-1.5 | LLaDA-1.5 · HumanEval 0-shot 256 | accuracy (%): 43.29 -> 36.59 | Hurt |
| Elastic-Cache with more accurate base model | Elastic-Cache · LLaDA-1.5 vs LLaDA GSM8K | speedup: 25.2x -> 45.1x (512 tokens) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 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 11 of 11 shared results
- dLLM-Cache — dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching (ICML 2026) · better on 1 of 1 shared result
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
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.94)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.93)
- EntropyCache: Decoded Token Entropy Guided KV Caching for Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.91)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.90)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 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.
