All articles tagged with #funsearch
Google DeepMind has successfully used its FunSearch method, which combines large language models with an automated evaluator, to solve an age-old mathematical problem known as the cap set problem. This marks the first time a long-standing scientific puzzle has been cracked using a large language model, according to a peer-reviewed study published in Nature.
Researchers at Google DeepMind have made what they claim to be the world's first scientific discovery using a large language model (LLM), suggesting that chatbot technology can generate information beyond human knowledge. The breakthrough emerged from DeepMind's investigation into whether LLMs, such as OpenAI's ChatGPT and Google's Bard, can produce new insights. DeepMind's FunSearch, which combines an LLM with an evaluator, was able to solve two puzzles in pure mathematics and computer science, generating new solutions that surpass those previously discovered by human mathematicians. The researchers believe this development could have transformative implications for computer science and algorithmic discovery.
DeepMind's AI system, FunSearch, has demonstrated its ability to generate new solutions to mathematical problems in the field of combinatorics, surpassing what was previously known by mathematicians and computer scientists. FunSearch utilizes large language models to create short computer programs that can generate solutions to specific mathematical problems. The system then checks the generated solutions against known ones, providing feedback to improve future iterations. The AI system was able to improve on the lower bound for a specific problem in combinatorics, showcasing the potential for human-machine collaboration in mathematics.
Researchers have introduced FunSearch, an evolutionary procedure that pairs a pre-trained Large Language Model (LLM) with a systematic evaluator to overcome confabulations and make new discoveries. By applying FunSearch to the cap set problem in extremal combinatorics, they have found new constructions of large cap sets, surpassing previous results. FunSearch has also been applied to the algorithmic problem of online bin packing, resulting in new heuristics that outperform existing baselines. This approach not only improves the effectiveness and scalability of LLM-based approaches but also produces more interpretable programs, enabling collaboration between domain experts and FunSearch for real-world applications.