EliteTable Boosts Control Policy Search with LLM Guidance
EliteTable leverages LLMs and a fixed‑capacity solution table to iteratively improve control policies, achieving notable score gains on multiple Gymnasium benchmarks.
EliteTable is a new program‑search algorithm that incorporates large language models (LLMs) to propose, edit, and remix high‑scoring solutions stored in a fixed‑capacity table. The approach was tested on six Gymnasium control tasks, running ten generations with fifty candidate slots per generation and ten independent search initializations per task.
Across all six tasks, median candidate scores increased from the first to the final generation, and median scores of the selected winners rose in five of the tasks. In the CartPole‑v1 environment, the selected winner hit the search‑score ceiling already in the first generation of every search run.
Four of the six tasks showed improvement on all ten held‑out endpoints, while BipedalWalker‑v3 and CarRacing‑v3 also improved, though they fell short of their predefined reporting targets. The results demonstrate EliteTable’s capacity to accelerate the discovery of effective executable control policies through LLM‑driven self‑improvement.
The study, posted on arXiv (cs.AI) on October 12, 2026, highlights the potential of combining LLM reasoning with structured solution tables to enhance reinforcement‑learning style searches without expanding model size.
Read also
Key Trends in LLMs Highlighted at WWC North America 2026
Simon Willison delivered a closing keynote at the WeAreDevelopers World Congress North America, summarizing 2026 LLM milestones.
OpenAI Unveils 722 Math Papers Solving Long‑Standing Problems
OpenAI announced a batch of 722 unpublished manuscripts that contain solutions to numerous enduring math challenges, grouped into 372 result families.
DEEPO Tackles Hallucination in Multimodal LLMs with Dual-Entropy RL
Researchers propose DEEPO, a dual‑entropy reinforcement learning approach designed to curb hallucinations in multimodal LLMs.
