Singular Proxies for Adaptive Caching in Diffusion Language Models

Wenhao Sun, Rong-Cheng Tu §, Yifu Ding, Zhao Jin, Jingyi Liao, Yongcheng Jing, Dacheng Tao § · 2026

Read Singular Proxies for Adaptive Caching in Diffusion Language Models by Wenhao Sun, Rong-Cheng Tu §, Yifu Ding, Zhao Jin, Jingyi Liao, Yongcheng Jing, Dacheng Tao § (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02544.txt.

Description

Wenhao Sun and 6 co-authors speed up diffusion language models by caching hidden states adaptively, in a 2026 arXiv paper. Diffusion language models generate text in arbitrary order, a flexible alternative to left-to-right generation. Their non-causal structure blocks the standard key-value caching trick, so every decoding step recomputes everything at full cost.

Existing caches update selected hidden states but rely on expensive token-wise heuristics and rigid, uniform budgets. SPA-Cache instead derives a low-dimensional singular proxy that flags update-critical tokens cheaply. An adaptive strategy then gives fewer updates to stable layers without hurting quality. Together these changes yielded up to eight times the throughput of uncached decoding. Speedups of two to four times over existing caching baselines were also reported. The work removes a central efficiency barrier for this model family.

Catalog record

CREATORWenhao Sun, Rong-Cheng Tu §, Yifu Ding, Zhao Jin, Jingyi Liao, Yongcheng Jing, Dacheng Tao §
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE61 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-02544

Browse the full library · About the Tower of Babel