FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution
FrugalEvo frames LLM-guided program search around cost per gain, not just final score.
The paper pairs a stronger, more expensive model for strategy exploration with a cheaper model for implementation and code refinement. It adds a cache-efficient evolution setup and evaluates progress with Budget-Aware AUC, tracking best score over cumulative LLM spend. Across 10 mathematical and systems optimization tasks, it reports higher BA-AUC on 9 and final quality matching or beating several baselines. On circle packing, the authors claim state-of-the-art results at $1.68 with GPT-5.6 Terra/Luna and $0.55 with GLM-5.3 plus Flash, versus roughly $50 for cited multi-agent baselines. HF Daily Papers' note
The paper pairs a stronger, more expensive model for strategy exploration with a cheaper model for implementation and code refinement. It adds a cache-efficient evolution setup and evaluates progress with Budget-Aware AUC, tracking best score over cumulative LLM spend. Across 10 mathematical and systems optimization tasks, it reports higher BA-AUC on 9 and final quality matching or beating several baselines. On circle packing, the authors claim state-of-the-art results at $1.68 with GPT-5.6 Terra/Luna and $0.55 with GLM-5.3 plus Flash, versus roughly $50 for cited multi-agent baselines. HF Daily Papers' note
score 5