Reproduce “Online Prompt Selection for Program Synthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yixuan Li, Lewis Frampton, Federico Mora, Elizabeth Polgreen
Researchers developed CYANEA, an online learning system for program synthesis that intelligently selects between symbolic solvers and various LLM-prompt combinations. This system enhanced the percentage of solved queries by 37.2% compared to the best single solver, approaching the theoretical optimum while also reducing computational cost and time.
Key insights
- A single-layer k-NN Multi-Armed Bandit configuration consistently outperformed a multi-layer hierarchical approach for solver and prompt selection.
- The system's custom reward functions effectively guided its strategy to prioritize different user objectives, such as solving speed, cost efficiency, or maximizing the number of solved queries.
- Dynamic budget allocation based on estimated solver performance improved efficiency, positively impacting both Par-2 scores and average solving costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.