Researchers isolate memorization from reasoning in AI neural networks
Researchers from Goodfire.ai have discovered that in AI language models like GPT-5, memorization and reasoning operate through separate neural pathways. By removing memorization pathways, models lost their ability to recite training data but retained logical reasoning skills. Surprisingly, arithmetic operations share neural pathways with memorization rather than reasoning, explaining AI models' struggles with math. This finding sheds light on how AI language models handle information and highlights the distinction between logical reasoning and mathematical reasoning in AI systems.
Ars Technica•