AdaStop Cuts AI Testing Costs via Dynamic Stopping
A new cost-benefit framework allows developers to stop labeling test data once the marginal value of finding faults drops below the cost of labor.
AdaStop optimizes the expensive process of validating deep neural networks by treating test selection as a financial decision. While traditional testing relies on a fixed budget that often results in either missed failures or wasted spend, this framework stops labeling the moment the marginal fault discovery rate falls below a specific cost-benefit threshold.
https://arxiv.org/abs/2607.05461
The system calculates a threshold based on the cost of labeling an input versus the value of discovering a fault. By estimating the discovery rate in real time, the framework removes the guesswork from budget allocation.
Experimental results demonstrate significant efficiency gains across various architectures and datasets. The approach discovered 65% to 84% of faults while utilizing only 9% to 31% of the total labeling budget.
This shift toward adaptive stopping reduces the human-labeling overhead required for quality assurance. It transforms AI validation from a fixed-cost exercise into a variable-cost process that scales with actual model risk.