FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution
FrugalEvo reports stronger optimization per dollar by splitting strategy search and code refinement across differently priced LLMs.
The paper introduces a cost-aware evolution loop: a higher-cost model explores strategies while a cheaper model implements and refines code. It adds cache-efficient prompting and a BA-AUC metric that tracks best-so-far score against cumulative LLM cost. Across 10 math and systems optimization tasks, it matches or beats several evolutionary baselines and posts higher BA-AUC on 9 of them. On circle packing, the authors report state-of-the-art results at $1.68 with GPT-5.6 Terra/Luna and $0.55 with GLM-5.3/Flash, versus about $50 average cost for cited multi-agent baselines. ArXiv · AI/CL/LG's note
The paper introduces a cost-aware evolution loop: a higher-cost model explores strategies while a cheaper model implements and refines code. It adds cache-efficient prompting and a BA-AUC metric that tracks best-so-far score against cumulative LLM cost. Across 10 math and systems optimization tasks, it matches or beats several evolutionary baselines and posts higher BA-AUC on 9 of them. On circle packing, the authors report state-of-the-art results at $1.68 with GPT-5.6 Terra/Luna and $0.55 with GLM-5.3/Flash, versus about $50 average cost for cited multi-agent baselines. ArXiv · AI/CL/LG's note
score 4