Saved in:
Bibliographic Details
Main Authors: Zuo, Bowen, Zhou, Dongruo, Zhu, Yinglun
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.21018
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908987859927040
author Zuo, Bowen
Zhou, Dongruo
Zhu, Yinglun
author_facet Zuo, Bowen
Zhou, Dongruo
Zhu, Yinglun
contents While scaling test-time compute can substantially improve model performance, existing approaches either rely on static compute allocation or sample from fixed generation distributions. In this work, we introduce a test-time compute allocation framework that jointly adapts where computation is spent and how generation is performed. Our method begins with a warm-up phase that identifies easy queries and assembles an initial pool of question-response pairs from the test set itself. An adaptive phase then concentrates further computation on unresolved queries while reshaping their generation distributions through evolving in-context demonstrations -- conditioning each generation on successful responses from semantically related queries rather than resampling from a fixed distribution. Experiments across math, coding, and reasoning benchmarks demonstrate that our approach consistently outperforms existing baselines while consuming substantially less inference-time compute.
format Preprint
id arxiv_https___arxiv_org_abs_2604_21018
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Adaptive Test-Time Compute Allocation with Evolving In-Context Demonstrations
Zuo, Bowen
Zhou, Dongruo
Zhu, Yinglun
Artificial Intelligence
While scaling test-time compute can substantially improve model performance, existing approaches either rely on static compute allocation or sample from fixed generation distributions. In this work, we introduce a test-time compute allocation framework that jointly adapts where computation is spent and how generation is performed. Our method begins with a warm-up phase that identifies easy queries and assembles an initial pool of question-response pairs from the test set itself. An adaptive phase then concentrates further computation on unresolved queries while reshaping their generation distributions through evolving in-context demonstrations -- conditioning each generation on successful responses from semantically related queries rather than resampling from a fixed distribution. Experiments across math, coding, and reasoning benchmarks demonstrate that our approach consistently outperforms existing baselines while consuming substantially less inference-time compute.
title Adaptive Test-Time Compute Allocation with Evolving In-Context Demonstrations
topic Artificial Intelligence
url https://arxiv.org/abs/2604.21018