AI research wiki
agentic-systems · 2026-10-06 · Tier 2

FrugalEvo: AlphaEvolve-style search scored per dollar

FrugalEvo: AlphaEvolve-style search scored per dollar

Source: HuggingFace Daily Papers, 2026-10-05 · arXiv 2610.03675 · raw: raw/huggingface/2026-10-05-frugalevo-towards-cost-aware-llm-guided-program-evolution.md

TL;DR. LLM-guided program evolution (AlphaEvolve, OpenEvolve and kin) is usually judged by gain after a fixed number of iterations. FrugalEvo judges gain per dollar. A strong, expensive model proposes strategies; a cheap model implements and refines the code. The harness and prompts are arranged so evolution steps share long prefixes, raising prompt-cache hits. A new metric, BA-AUC, is the area under the best-so-far score curve over cumulative LLM spend up to a budget. Across 10 math and systems optimization tasks it matches or beats OpenEvolve, ShinkaEvolve, AdaEvolve and EvoX on final quality and wins BA-AUC on 9. On circle packing it sets a new best for $1.68 (GPT-5.6 Terra + Luna) or $0.55 (GLM-5.3 + Flash), against multi-agent systems like CORAL and SwarmResearch at about $50.

Expensive model plans, cheap model types, the cache does the rest
Prefix-shared prompts turn most of each step into cache reads.
flowchart LR
  P["Program pool<br/><small>best-so-far</small>"] --> X["Explorer LLM<br/><small>strong, pricey</small>"]
  X -->|strategy| I["Implementer LLM<br/><small>cheap, iterates</small>"]
  I --> E["Evaluator<br/><small>task score</small>"]
  E --> P
  C["Shared prefix<br/><small>cache hits</small>"] --> X
  C --> I
  E --> M["BA-AUC<br/><small>score per dollar</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  class P,C input
  class X,I core
  class E loop
  class M exit
  linkStyle 3 stroke:#f08c00,stroke-width:2px
Blue are the pool and cached prefix, purple are the two models, amber is the evaluate-and-loop step, green is the cost-aware metric.

How it relates to prior wiki pages

  • A routing paper in disguise: plan-with-big, execute-with-small is the same split as cascade routing on the LLM routing page, applied inside a search loop.
  • Confirms the 10-05 cost-unit thread: Beyond Token Savings (10-05) argued dollars and seconds per finished task, not tokens, are the right scoreboard. BA-AUC is that scoreboard for evolutionary search.
  • Counter-evidence to swarm spend: beating ~$50 multi-agent systems for under $2 matches Ord's finding (see swarm scaling) that swarms buy speed at a token premium.
  • Cache-aware prompt layout echoes Kepler (10-04: 97.37% cache reads in an agent benchmark bill).

Gaps

  • Circle packing and similar tasks have cheap evaluators; costs will look different where evaluation dominates.
  • The explorer/implementer split is fixed, not learned.

Related

Agent harness engineering · Self-evolving agents · Daily digest 2026-10-06