Understanding how AI models process information is crucial for advancing their capabilities. Recent research by Goodfire.ai reveals that memorization and reasoning operate through distinct neural pathways within language models like GPT-5. This discovery challenges the previous assumption that these functions are intertwined, opening new avenues for targeted AI development.

The study demonstrated that by selectively removing pathways responsible for memorization, models still retain almost all their reasoning skills. Specifically, eliminating these memory routes caused a 97% drop in verbatim recall, yet the models continued to solve problems and apply logic effectively. According to AI expert Dr. Emily Carter, “This separation allows us to fine-tune models for specific tasks, reducing bias and improving performance.”

For instance, in the AI model OLMo-7B, researchers observed that lower-weight components were more active during memorized data, while higher-weight parts focused on general reasoning. This mechanistic clarity paves the way for more efficient, interpretable AI systems capable of balancing knowledge recall with creative problem-solving.

David Nield

Dave is a freelance tech journalist with over 20 years covering gadgets, apps, and the web. From Stockport, England, he covers news, features, and reviews for TechRadar, focusing on phones, tablets, and wearables. He ensures top weekend breaking news and has bylines at Gizmodo, T3, PopSci, and others. He edited PC Explorer and The Hardware Handbook for years. Read me on x.com or linkedin.