EliteTable is a self‑improving program‑search algorithm guided by large language models (LLMs). It keeps a fixed‑capacity table of high‑scoring solutions, which are used to propose new candidates, edit existing ones, and remix multiple solutions.
The method was evaluated on six Gymnasium control tasks. Each task ran ten generations with fifty candidate slots per generation, and ten independent searches were started for each task. Candidate‑score medians increased from the first to the final generation in all six tasks, and winner‑score medians rose in five of them. In CartPole‑v1, the selected winner reached the search‑score ceiling in the first generation of all ten searches. Four tasks showed improvement across all ten held‑out endpoints; BipedalWalker‑v3 and CarRacing‑v3 improved without meeting the reporting targets. All ten Acrobot‑v1 searches declined on held‑out episodes, yielding a negative median change.
These results demonstrate that performance can improve during program search and that gains of the selected winner can transfer to new episodes in some tasks. Further comparative experiments and ablations are required to establish EliteTable’s relative performance and the contribution of each component.
Review