FrugalEvo: cost-aware LLM evolution pairs expensive strategist with cheap coder, cutting per-iteration cost ~10x
PangWeiKoh · x · 2026-10-15
Researchers from NUS, UW, Princeton and Stanford introduce FrugalEvo, an open-source, cost-aware LLM-guided evolutionary optimization framework.
- Idea: unlike AlphaEvolve-style methods measured over fixed iterations, FrugalEvo maximizes gain per unit cost — a stronger, pricier LLM explores solution strategies while a cheaper LLM implements and iteratively refines the code.
- Cache efficiency: the harness and prompts are designed to maximize prefix sharing across evolution steps for better cache reuse.
- New metric: Budget-Aware Area Under the Curve (BA-AUC) measures solution quality under a fixed cost budget.
- Results: on circle packing, FrugalEvo with GLM 5.3 + GLM 5.3 Flash costs $0.0097 per iteration vs $0.021–0.043 for baselines, achieving better solutions within the same $1 budget.
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23