AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Singular Proxies for Adaptive Caching in Diffusion Language Models

Wenhao Sun, Rong-Cheng Tu, Yifu Ding, Zhao Jin, Jingyi Liao, Yongcheng Jing, Dacheng Tao

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Distributed LLM inference systems

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 changedWhereMeasuredEffect
replace Value proxy with singular proxy (rank 128) (ablation)SPA-Cache · GSM8K, LLaDA-8BTPS: 165 -> 179Improved
replace Value proxy with singular proxy (rank 128) (ablation)SPA-Cache · GSM8K, LLaDA-8Baccuracy (%): 78.59 -> 78.23No clear effect
add adaptive layer-wise budget allocation (ablation)SPA-Cache · GSM8K, LLaDA-8BTPS: 179 -> 189Improved
use uniform 16% update ratio instead of adaptive allocation (ablation)SPA-Cache · GSM8K, LLaDA-8Baccuracy (%): 78.23% to 75.56%Hurt
lower proxy rank from 512 to 32 (ablation)SPA-Cache · GSM8K, LLaDA-8Baccuracy (%): 78.57 -> 76.71 accuracy, 172.56 -> 182.66 TPSMixed
use Attention output as identifier instead of Value (ablation)SPA-Cache · identification accuracyidentification accuracyHurt
add SPA-Cache to LLaDA-8BLLaDA-8B-Instruct · MBPPTPS: 5.75 -> 46.12Improved
add SPA-Cache to LLaDA-8BLLaDA-8B-Instruct · GSM8KTPS: 29.67 -> 190.73Improved
add SPA-Cache to LLaDA-8BLLaDA-8B-Instruct · GSM8Kaccuracy (%): 78.62 -> 78.24No clear effect
add Fast-dLLM instead of SPA-Cache to Dream on HumanEval accuracyFast-dLLM · HumanEval, Dream-7Baccuracy (%): 58.54 -> 53.05Hurt
combine SPA-Cache with parallel decodingSPA-Cache · BBH, LLaDA-8BTPS: 108.31 -> 693.96Improved
use SPA-Cache instead of Fast-dLLM with parallel decodingSPA-Cache · BBH, LLaDA-8BTPS: 301.33 -> 693.96Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy