AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery
The paper’s claim is that an agent can improve algorithm discovery by editing code from runtime feedback and reusing learned design skills.
AlgoEvo replaces fixed search pipelines with an autonomous loop that inspects, diagnoses, and modifies executable code. The framework separates reusable design knowledge into a skill hub, so one workflow can handle single-objective, multi-objective, and multi-component design tasks. It also stores search experience in a hierarchy, using task-level trajectories and cross-task patterns to guide later exploration. Across six benchmark tasks, the authors report that AlgoEvo matched or beat specialized methods while using fewer evaluations and tokens. ArXiv · AI/CL/LG's note
AlgoEvo replaces fixed search pipelines with an autonomous loop that inspects, diagnoses, and modifies executable code. The framework separates reusable design knowledge into a skill hub, so one workflow can handle single-objective, multi-objective, and multi-component design tasks. It also stores search experience in a hierarchy, using task-level trajectories and cross-task patterns to guide later exploration. Across six benchmark tasks, the authors report that AlgoEvo matched or beat specialized methods while using fewer evaluations and tokens. ArXiv · AI/CL/LG's note
score 5