Summary
AI systems are often improved by using more resources, but just because a bigger budget leads to better results doesn’t mean it’s clear how or where to spend those resources best. The authors looked critically at studies that try to figure this out and found that many comparisons make some common mistakes, like measuring success before the system picks an answer or ignoring some costs. They offer a new way to think about performance that accounts for all resources and decisions involved. This helps clarify which resource choices we can trust from the data, and which questions still need answers.
What this means in practice
- •For machine learning engineers: Understand when reported improvements justify shifting computing or data resources to optimize AI system performance.
- •For product managers for ai tools: Guide decisions on investing in pretraining, inference time, or additional data retrieval to improve deployed AI systems effectively.
A survey. It maps existing work.
Abstract
AI scaling studies increasingly evaluate systems that combine a pretrained model with retrieval, search, verification, tools, and interaction. Yet a higher score under a larger budget does not by itself show where additional resources are best spent. This critical integrative review asks when a reported scaling result supports a resource-allocation decision. It compares evidence across pretraining, test-time computation, retrieval, and agent evaluation, distinguishing the performance of a tested procedure from the best performance achievable under a resource limit. The synthesis shows that three mismatches recur across this evidence: success counted before an answer is chosen, information a deployed system will not have, and costs left out of the comparison. A capability surface expresses performance as a function of budgets, mechanisms, and available information. Worked analytical examples show how the evaluation metric, deployment volume, selection rule, and stopping policy can alter an allocation conclusion. A resource envelope provides a structured record of the task, development and run-time resources, information access, and procedure behind a reported score. Its application to a published comparison illustrates which conclusions the evidence supports and which deployment questions remain unresolved. The resulting framework specifies the comparisons needed to choose among feasible systems and motivates experiments on the transfer of allocation rules across tasks and operating conditions. It does not propose a universal scaling law or infer general intelligence from benchmark gains.