FunSearch uses an evolutionary method powered by LLMs to develop high-scoring ideas, which are expressed as computer programs. This iterative process allows for the automatic evaluation and improvement of these programs. The method also outputs programs that reveal how its solutions are constructed, providing insights for further scientific discovery. The article suggests that FunSearch could be used to tackle a variety of scientific and engineering challenges in the future.
Key takeaways:
- A new method called FunSearch has been introduced, which uses Large Language Models (LLMs) to search for new solutions in mathematics and computer science. It pairs a pre-trained LLM with an automated evaluator to guard against incorrect information and evolve initial solutions into new knowledge.
- FunSearch has made the first discoveries in open problems in mathematical sciences using LLMs. It discovered new solutions for the cap set problem, a longstanding open problem in mathematics, and more effective algorithms for the “bin-packing” problem.
- FunSearch generates programs that reveal how its solutions are constructed, rather than just what the solutions are. This makes it a powerful scientific tool as it can inspire further insights in the scientists who use it.
- FunSearch is not only useful for mathematical discoveries but also for revealing potentially impactful solutions to important real-world problems. It is expected to become a common practice for generating effective and tailored algorithms for many problems in science and industry.