[Paper] SeaEvo: Advancing Algorithm Discovery with Strategy Space Evolution

Source: arXiv - 2604.24372v1

Overview

The paper SeaEvo proposes a new way to make large‑language‑model (LLM)‑driven evolutionary search more effective by giving each candidate program a persistent natural‑language description of its underlying strategy. By treating these strategy texts as first‑class citizens in the evolutionary population, the system can reason about, cluster, and navigate the “strategy space” rather than just the raw code and scalar fitness scores. The authors show that this approach yields consistent performance gains on a variety of algorithm‑discovery benchmarks.

Key Contributions

• Strategy‑Space Layer: Introduces a modular component that attaches a natural‑language strategy description to every program in the population.
• Strategy Articulation: Turns mutation into a three‑step “diagnose → direct → implement” workflow, letting the LLM explicitly reason about what to change and why.
• Stratified Experience Retrieval: Organizes the archive into clusters of similar strategies and selects mutation inspirations based on complementary behavior rather than raw similarity.
• Strategic Landscape Navigation: Periodically summarizes the current strategy landscape (effective, saturated, under‑explored) and feeds this high‑level guidance back into the mutation process.
• Empirical Validation: Demonstrates improvements across three benchmark families (mathematical algorithm discovery, systems optimization, and agent‑scaffold tasks), with up to 21 % relative gain on open‑ended system‑optimization problems.

Methodology

1. Population Augmentation – Each program candidate is paired with a short natural‑language paragraph that describes the algorithmic idea or design pattern it embodies (e.g., “divide‑and‑conquer search on sorted arrays”).
2. Mutation via Articulation – When mutating a candidate, the LLM first diagnoses the current strategy’s strengths/weaknesses, then directs a concrete change (e.g., “replace linear scan with binary search”), and finally implements the change to produce new code.
3. Strategy Clustering – The archive is periodically re‑clustered using embeddings of the strategy texts. Clusters represent families of related ideas.
4. Experience Retrieval – For a target candidate, the system pulls examples from complementary clusters (those that solve different sub‑problems) to inspire mutations, encouraging diversity.
5. Landscape Navigation – A lightweight summarizer runs every few generations, producing a meta‑strategy report (e.g., “binary‑search families are saturated; explore probabilistic pruning”). This report is fed back as a prompt to guide the next round of articulations.

All steps rely on off‑the‑shelf LLM APIs (e.g., GPT‑4) and standard evolutionary operators (selection, crossover). The novelty lies in persisting the strategy description and using it as an evolutionary signal.

Results & Findings

• Higher success rates were observed especially on open‑ended tasks where many low‑fitness but strategically promising programs exist.
• Reduced premature convergence: The strategy clusters prevented the population from collapsing onto a single syntactic variant of an idea.
• Improved sample efficiency: Fewer generations were needed to reach a given performance threshold (≈15 % fewer evaluations on average).

Practical Implications

• More Robust Auto‑ML Pipelines – By exposing the “why” behind each model‑training script, developers can let the system preserve promising but currently under‑performing pipelines for later refinement.
• Knowledge‑Base Construction – The persistent strategy texts can be harvested into a searchable repository of algorithmic patterns, enabling future projects to reuse high‑level designs without re‑deriving them.
• Debugging & Auditing – When a generated program fails, the accompanying strategy description offers an immediate human‑readable hypothesis, speeding up root‑cause analysis.
• Hybrid Human‑AI Collaboration – Engineers can inject or edit strategy descriptions directly, steering the evolutionary search toward domains of interest without writing low‑level code.
• Scalable Evolutionary Search – The clustering and retrieval mechanisms scale well with large archives, making SeaEvo suitable for cloud‑based, distributed search services that need to manage millions of candidate programs.

Limitations & Future Work

• Reliance on LLM Quality – The approach assumes the LLM can generate accurate, concise strategy descriptions; noisy or overly generic texts can degrade clustering.
• Computational Overhead – Periodic clustering and summarization add extra compute cycles, which may be non‑trivial for very large populations.
• Domain Transferability – Experiments focused on algorithmic and systems tasks; applying SeaEvo to domains like UI generation or hardware design may require domain‑specific prompting strategies.
• Future Directions – The authors suggest (1) integrating retrieval‑augmented generation to ground strategies in external documentation, (2) exploring hierarchical strategy representations (sub‑strategies), and (3) formalizing a “strategy‑level fitness” metric to better balance exploration vs. exploitation.

Authors

• Sichun Luo
• Yi Huang
• Haochen Luo
• Fengyuan Liu
• Guanzhi Deng
• Lei Li
• Qinghua Yao
• Zefa Hu
• Junlan Feng
• Qi Liu

Paper Information

• arXiv ID: 2604.24372v1
• Categories: cs.CL, cs.AI, cs.NE
• Published: April 27, 2026
• PDF: Download PDF