Researchers isolate memorization from reasoning in AI neural networks

Researchers at AI startup Goodfire.ai have discovered that memorization and reasoning operate through completely separate neural pathways in AI language models. When they surgically removed memorization circuits from models, the systems lost 97 percent of their ability to recite training data but kept nearly all logical reasoning abilities intact. Surprisingly, arithmetic operations share pathways with memorization rather than reasoning, causing math performance to plummet when memorization was removed, which may explain why AI models struggle with mathematics without external tools.