Many demanding puzzles require finding a binary code that satisfies many equations (Market Split) or the largest set of nodes with no direct links (maximum independent set). As these puzzles grow, they become extremely hard, and some benchmark instances had remained unsolved.
Two computer agents, each powered by a large language model, were used to design new algorithms. A human researcher guided the experiments, using diagnostic tests to steer the search.
The resulting algorithm found the first conventional-computer solutions to 18 previously open instances and produced test networks where the best existing solvers fell 5 to 15 nodes short of the known optimum. These benchmarks reveal when pure search fails and help design better algorithms.
