Test-time Scaling of LLMs: A Survey from A Subproblem Structure Perspective

TL;DR

This survey reviews test-time scaling techniques for large language models, emphasizing subproblem decomposition and organization, unifying diverse methods like Chain-of-Thought and Tree-of-Thought, and discussing their strengths and future directions.

Contribution

It introduces a unified perspective on test-time scaling methods based on subproblem structures, connecting various approaches under a common framework.

Findings

Unifies diverse test-time scaling approaches through subproblem structure analysis

Highlights strengths and weaknesses of existing techniques

Suggests promising future research directions

Abstract

With this paper, we survey techniques for improving the predictive accuracy of pretrained large language models by allocating additional compute at inference time. In categorizing test-time scaling methods, we place special emphasis on how a problem is decomposed into subproblems and on the topological organization of these subproblems whether sequential, parallel, or tree-structured. This perspective allows us to unify diverse approaches such as Chain-of-Thought, Branch-Solve-Merge, and Tree-of-Thought under a common lens. We further synthesize existing analyses of these techniques, highlighting their respective strengths and weaknesses, and conclude by outlining promising directions for future research