Test-Time Improvement in Large Language Model Systems: Resource Control, Evaluation, and Recovery
Yiming Tang , Ziyu Liu , Zihan Chen , Jinghan Yang , Liuyu Xiang , Zhenbo Xu , Zhaofeng He
Large language model (LLM) systems can be improved after training through mechanisms that alter their problem-solving behavior during inference or deployment. These mechanisms span prompting, retrieval augmentation, test-time adaptation, reasoning-time search, verification, memory, tool use, and agentic workflows. Although often studied as separate research directions, they are increasingly combined in modern LLM systems. This survey examines how these mechanisms can be compared and selected for different resource conditions, task complexities, and application scenarios. Taking their system-level implementation as its analytical object, it uses a lightweight Locus × Source map to organize where an intervention acts and where its enabling signal originates. Temporal persistence and multidimensional cost provide auxiliary descriptors. Resource control provides a cross-mechanism analytical perspective on allocation, stopping, state updating and retention, verification, and recovery. The survey further synthesizes evaluation, failure propagation, and recovery requirements, including multidimensional cost, evaluator reliability, protocol sensitivity, temporal error propagation, and persistent-state contamination. Together, these analyses connect the choice of an improvement strategy to the resources it uses, the duration of its effects, and the evaluation and failure-handling requirements of deployment. Additional resources alone do not guarantee improvement, so their value must be assessed in relation to task demands and how they are controlled.
Large language model systems / Test-time improvement / Test-time improvement mechanisms / Resource control / Multidimensional evaluation / Failure propagation and recovery
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(4350KB)
