Neural Networks Can Suddenly Grok After Plateau

Researchers trained a tiny neural network on modular addition in 2022. At 1,000 steps it scored perfectly on training data but only about 10 percent on new problems. By 20,000 steps, with no new data, it scored close to 100 percent on both. A 2023 follow-up found the network had learned circular number representations and rotation, essentially rediscovering trigonometry.