Large language models are now common tools for preparing data before it feeds into decisions—removing noise, flagging mislabeled examples, and filling gaps. But not all LLMs are equal. A more capable model may clean data more reliably, yet its higher cost per call changes the economics of the pipeline. The arXiv paper "When is a better LLM worth paying for? Stochastic Optimization with Switching Mixed-Fidelity Oracles" formalizes this as a decision problem.

The authors treat data curation as a stochastic optimization task in which an algorithm can query oracles of different fidelity—cheap but noisy, or expensive but accurate. The central question is when to switch from a low-fidelity LLM to a high-fidelity one, given that the value of a cleaner dataset depends on the downstream decision it supports. The proposed approach appears to model this switching dynamically, rather than committing to a single model for the entire pipeline.

Because the abstract is truncated, the full details of the method and its empirical results are not available here. Still, the framing is useful: it shifts the conversation from "which LLM is best" to "which LLM is best for this step, at this cost." That is a more practical question for real deployments, where budgets are finite and not every data error carries the same consequence.