AI systems that search for better programs can burn through model calls as they repeatedly propose code, test it and try again. FrugalEvo, a new framework from researchers at the National University of Singapore, the University of Washington, Princeton and Stanford, splits that cycle between a stronger model that proposes strategies and a cheaper one that writes and refines the code.
For each task, an evaluator runs candidate programs and scores them. FrugalEvo uses those results to decide which proposed strategy to pursue next, while keeping earlier programs and their feedback in a search memory. The researchers also arrange repeated prompt content so supported APIs can reuse cached input, reducing some repeated-token costs.
The paper evaluates the approach on 20 optimization tasks, from mathematical problems to systems work such as placing AI models across GPUs, plus algorithmic challenges. To measure progress within a spending limit, the authors introduce Budget-Aware Area Under the Curve, which tracks the best score found as model costs accumulate. FrugalEvo had a higher score on this measure in nine of the ten math and systems tasks, and the highest average on the ten algorithmic tasks, according to the authors’ tests against existing search methods.
The results apply to problems where candidate solutions can be generated and scored in code. The authors say the approach does not directly transfer to work requiring physical experiments, such as wet-lab research, and that its strategy instructions were hand-written for these tasks. They also caution that the generated solutions have not been validated by human experts. The study therefore supports a narrower claim: splitting planning and coding between models can improve benchmark search progress under a fixed model-cost budget.
Sources
- FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution (arXiv preprint, Oct. 2, 2026).
- Official FrugalEvo implementation (Apache-2.0 license).
Image credit: FrugalEvo authors, Figure 2 (CC BY 4.0).