UniScale : Adaptive Unified Inference Scaling via Online Joint Optimization of Model Routing and Test-Time Scaling

Kaiyu Huang, Xingyu Wang, Mingze Kong, Zhubo Shi, Yuqian Hou, Hong Xu, Zhongxiang Dai, Minchen Yu, Qingjiang Shi · 2026

Read UniScale : Adaptive Unified Inference Scaling via Online Joint Optimization of Model Routing and Test-Time Scaling by Kaiyu Huang, Xingyu Wang, Mingze Kong, Zhubo Shi, Yuqian Hou, Hong Xu, Zhongxiang Dai, Minchen Yu, Qingjiang Shi (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-2605-30898.txt.

Description

Huang and colleagues unify two ways of scaling language-model inference that are tuned apart. Model routing swaps among varied models to fit request difficulty, but gives coarse jumps; test-time scaling tunes compute inside one model, but hits capacity ceilings and diminishing returns. Their UniScale framework joins routing and test-time scaling in one optimization space and learns policies with a contextual bandit method. Efficiency-aware learning and cost modeling keep search stable over action spaces. Tests show a finer quality-versus-cost trade-off across shifting settings.

Catalog record

CREATORKaiyu Huang, Xingyu Wang, Mingze Kong, Zhubo Shi, Yuqian Hou, Hong Xu, Zhongxiang Dai, Minchen Yu, Qingjiang Shi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE92075
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.30898

Browse the full library · About the Tower of Babel