EliteTable: Self-Improving Program Search Algorithm Guided by Large Language Models Released
Summary:EliteTable is a self-improving program-search algorithm guided by large language models (LLMs), utilizing a fixed-capacity table of high-scoring solutions to guide new proposals, edits of existing solutions, and remixes of several solutions. Evaluated across six Gymnasium control tasks, each with ten generations and fifty candidate slots per generation, the results demonstrate improvements in candidate-score medians across all tasks and selected-winner search-score medians in five tasks. Notably
Technical Mechanism Analysis
The core mechanism of EliteTable is based on the guidance of large language models (LLMs) and involves the following steps to achieve self-improvement in program search:
- High-Score Solutions Table: EliteTable maintains a fixed-capacity table of high-scoring solutions to store and update the current optimal program solutions.
- New Proposal Generation: New program proposals are generated using LLMs, which are optimized and adjusted based on the contents of the high-score solutions table.
- Editing Existing Solutions: Existing solutions are edited to adapt to new task requirements or improve their performance.
- Mixing Solutions: Multiple solutions are mixed to generate new, better programs.
Engineering Trade-offs and Performance
Experiments on six Gymnasium control tasks show that EliteTable can continuously improve the quality of solutions during program search. However, the degree of improvement varies across different tasks:
- CartPole-v1: Reached the search-score ceiling in the first generation across all searches, indicating excellent adaptability to EliteTable.
- BipedalWalker-v3 and CarRacing-v3: Showed improvements but did not reach their targets, suggesting these tasks pose higher challenges for EliteTable.
- Acrobot-v1: Some search results declined, indicating that EliteTable may need further optimization for certain complex tasks.
Developer Implementation and Deployment Recommendations
EliteTable offers a new approach to program search, particularly suitable for tasks that require rapid iteration and optimization. Developers can refer to the following recommendations for implementation:
- Task Adaptability Assessment: Before applying EliteTable to a specific task, it is recommended to conduct a thorough adaptability assessment to ensure it can effectively solve the target problem.
- Adjusting Solutions Table Capacity: Adjust the capacity of the high-score solutions table according to task complexity and resource constraints to balance search efficiency and storage overhead.
- Optimizing Mixing Strategies: Explore the impact of different mixing strategies on solution quality to find the optimal mixing scheme.
Conclusion
EliteTable demonstrates the potential of self-improving program search algorithms guided by large language models. Although there is still room for improvement in certain complex tasks, its overall performance is impressive, providing new research directions for the AI program generation and optimization field.
— END —Source: ArXiv AI (cs.AI) (2026-10-12)
Tags: #Large Language Models #Program Search #Self-Improving Algorithms #Gymnasium Control Tasks #AI Optimization
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments