FrugalEvo: Towards Cost-Aware LLM-Guided Program Evolution
FrugalEvo pairs a strong strategist LLM with a cheap implementer, achieving state-of-the-art circle packing for $1.68 and beating costlier baselines.
FrugalEvo is a cost-aware LLM-guided evolutionary optimization framework where an expensive LLM explores solution strategies while a cheaper LLM implements and refines the code, with cache-efficient prompt design maximizing prefix sharing. The paper introduces Budget-Aware AUC (BA-AUC) to measure solution quality per cumulative LLM cost. Across 10 mathematical and systems tasks it matches or surpasses OpenEvolve, ShinkaEvolve, AdaEvolve, and EvoX in final quality and wins BA-AUC on 9 tasks. On circle packing it reaches new state-of-the-art with GPT-5.6 Terra/Luna for $1.68 and GLM-5.3 (plus Flash variant) for $0.55, versus roughly $50 for multi-agent baselines like CORAL and SwarmResearch.
- Strong LLM plans strategies; cheap LLM implements and refines code
- New BA-AUC metric scores best-so-far quality against cumulative LLM cost
- Beats OpenEvolve, ShinkaEvolve, AdaEvolve, EvoX on 10 optimization tasks
- Circle-packing SOTA for $1.68 (GPT-5.6) or $0.55 (GLM-5.3)
- Multi-agent baselines CORAL and SwarmResearch cost around $50 on average
- AI models
- GLM-5.3
From the original article
1 min read at arxiv.orgLLM-guided evolutionary methods, such as AlphaEvolve, have emerged as powerful approaches for challenging computational optimization problems, such as circle packing. However, prior work typically optimizes performance gain over a fixed number of iterations. We argue that practical optimization should maximize gain per unit cost.
By Hui Chen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.