aiqbench analysis
ExpEvolve is an arXiv preprint that adds idea-level abstractions to LLM-driven evolutionary coding agents, in the vein of AlphaEvolve, so that search can reuse high-level factors that explain score differences rather than relying only on surface-level program mutations. The reported gains are up to 5.0x in sample efficiency across 11 mathematical tasks, with some transfer to other LLMs and six MLS-Bench machine learning tasks. These are benchmark results from a single research group, so readers should look for independent replication and check how the gains hold up under different compute budgets.
ExpEvolve: Structuring and Reusing Experience for Open-Ended Discovery
arXiv:2610.29303v1 Announce Type: new 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.
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.
Machine Learning (cs.LG); Computation and Language (cs.CL); Machine Learning (stat.ML)
arXiv:2610.29303 [cs.LG]
(or arXiv:2610.29303v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.29303 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 290 words
Related coverage
Tree-of-Experience: A Structured Experience-Management Solution for Self-Evolving Agents under Low-Repetition and Implicit-Reward Environments
arXiv · Jun 8, 2026
Trace Only What You Need: Structure-Aware On-Demand Hypergraph Memory for Long-Document Question Answering
arXiv · Aug 7, 2026
Trace Only What You Need: Structure-Aware On-Demand Hypergraph Memory for Long-Document Question Answering
arXiv · Jun 10, 2026
DeepInstructor: An Agentic AI Instructor for Experience-Driven Idea Evaluation
arXiv · Sep 22, 2026