Newest papers

ExpEvolve: Structuring and Reusing Experience for Open-Ended Discovery

Yongqiang Chen, Chenxi Liu, Deyu Zou, Zhenhao Chen, Tongliang Liu, Bo Han, Kun Zhang · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29303 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Los agentes de codificación evolutiva como AlphaEvolve son uno de los éxitos notables en el uso de Grandes Modelos de Idiomas (LLM) para construir científicos de IA. Estos agentes abordan problemas científicos abiertos mejorando e evolucionando programas iterativamente, aprovechando los conocimientos previos y las capacidades de razonamiento de los LLM. A pesar de su éxito, las mutaciones sobre los programas son principalmente de bajo nivel en heurística superficial, sin una unidad persistente de control. En contraste, la ciencia se compone a través de ideas. En este trabajo, proponemos ExpEvolve, que inyecta la extracción de nivel de idea y el control en la evolución: descubre y aprovecha los factores de alto nivel de los programas que explican las diferencias entre los diferentes programas. Caracterizamos las condiciones teór

Machine translation (Meta M2M100); the English original follows.

Abstract

Evolutionary coding agents such as AlphaEvolve are one of the notable successes in using Large Language Models (LLMs) to build AI Scientists. These agents tackle open-ended scientific problems by iteratively improving and evolving programs, leveraging the prior knowledge and reasoning capabilities of LLMs. Despite their success, mutations over programs are mostly low-level upon surface heuristics, without a persistent unit of control. In contrast, science compounds through ideas. In this work, we propose ExpEvolve, which injects idea-level extraction and control into evolution: it discovers and leverages the high-level factors from programs that explain scoring differences among different programs. We characterize theoretical conditions on when discovery with idea factors improves sample efficiency. We also build a practical framework to show that the extracted high-level idea factors are helpful for guiding exploitation from experience and exploration when discovery comes to a plateau. Across 11 open-ended mathematical tasks, ExpEvolve significantly improves the final performance and sample efficiency up to 5.0x. The improvements also transfer to 4 other LLMs and to 6 machine learning tasks from MLS-Bench. Analysis shows that ExpEvolve can discover idea-level factors to provide better exploration and exploitation for evolutionary discovery.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Reuse, Structuring, Artificial intelligence, Coding (social sciences), Machine learning

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange