FrugalEvo hits circle-packing SOTA for $1.68, undercutting ~$50 multi-agent baselines
Hui Chen · hf · 2026-10-05
Researchers propose FrugalEvo, a cost-aware LLM-guided program evolution framework arguing optimization should maximize gain per unit cost, not performance at a fixed iteration count.
Design:
- A stronger, pricier LLM explores solution strategies; a cheaper LLM implements and iteratively refines the code;
- Cache-efficient evolution: harness and prompts maximize prefix sharing across steps;
- New metric BA-AUC: area under the best-so-far score curve over cumulative LLM cost within a fixed budget.
Results:
- Matches or beats OpenEvolve, ShinkaEvolve, AdaEvolve, and EvoX on 10 math/systems tasks, with higher BA-AUC on 9;
- Higher average performance on 10 ALE-Bench-Lite algorithmic tasks;
- New circle-packing SOTA with GPT-5.6 Terra/Luna for $1.68 and GLM-5.3 (+Flash) for $0.55, vs $50 average for multi-agent baselines like CORAL and SwarmResearch.
More from Research
- First large-scale 3B/8B continuous diffusion LMs match pass@1 and beat pass@k vs masked dLMs — ArashVahdat · 2026-10-06
- AggAgent at COLM: Treats Parallel Agent Trajectories as an Environment for Long-Horizon Tasks — xiye_nlp · 2026-10-06
- New paper: LLM verbalized probabilities directly track internal uncertainty, not just training data — sineadwilliamso · 2026-10-06
- Nature Biotech paper: "single-cell models don't beat baselines" was a measurement problem — BoWang87 · 2026-10-06
- 250+ AI experts broadly agree: international body with pre-release power best cuts risk — robseamans · 2026-10-06
- Negative result: state-cosine adaptive depth stopping rule hurts both loss and passes — BlackHC · 2026-10-06