![[2602.20133] AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.20133] AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization
Summary
AdaEvolve introduces a novel framework for optimizing large language model-driven evolution, addressing inefficiencies in resource allocation during automated program generation.
Why It Matters
This research is significant as it enhances the efficiency of LLMs in evolutionary computing, potentially leading to breakthroughs in automated program generation and optimization. By addressing the limitations of static scheduling, AdaEvolve could improve resource utilization and solution quality across various optimization problems.
Key Takeaways
- AdaEvolve reformulates LLM-driven evolution as a hierarchical adaptive optimization problem.
- The framework utilizes an accumulated improvement signal for decision-making across multiple levels.
- AdaEvolve demonstrates superior performance over existing baselines in 185 optimization problems.
Computer Science > Neural and Evolutionary Computing arXiv:2602.20133 (cs) [Submitted on 23 Feb 2026] Title:AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization Authors:Mert Cemri, Shubham Agrawal, Akshat Gupta, Shu Liu, Audrey Cheng, Qiuyang Mang, Ashwin Naren, Lutfi Eren Erdogan, Koushik Sen, Matei Zaharia, Alex Dimakis, Ion Stoica View a PDF of the paper titled AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization, by Mert Cemri and 11 other authors View PDF HTML (experimental) Abstract:The paradigm of automated program generation is shifting from one-shot generation to inference-time search, where Large Language Models (LLMs) function as semantic mutation operators within evolutionary loops. While effective, these systems are currently governed by static schedules that fail to account for the non-stationary dynamics of the search process. This rigidity results in substantial computational waste, as resources are indiscriminately allocated to stagnating populations while promising frontiers remain under-exploited. We introduce AdaEvolve, a framework that reformulates LLM-driven evolution as a hierarchical adaptive optimization problem. AdaEvolve uses an "accumulated improvement signal" to unify decisions across three levels: Local Adaptation, which dynamically modulates the exploration intensity within a population of solution candidates; Global Adaptation, which routes the global resource budget via bandit-based scheduling across different solution candidate pop...