Discovering heuristics in a complex SAT solver with large language models
Researchers have developed AutoModSAT, a new framework that leverages large language models (LLMs) to automatically optimize complex SAT solvers. This approach combines an LLM-compatible modular solver design with unsupervised prompt optimization and an evolutionary algorithm. Experiments show that AutoModSAT significantly improves performance, achieving a 40% gain over the baseline solver and a 30% gain over state-of-the-art solvers, demonstrating LLMs' potential in heuristic discovery for optimization tasks. AI
IMPACT Demonstrates LLMs can optimize complex computational tools, potentially accelerating research and development in various fields.