Study Finds AI Models Store Memories and Logic in Different Neural Regions

Why It Matters

The finding reveals a tractable path to edit AI models for privacy, copyright compliance, and safety while retaining their utility, and it explains why current models struggle with math, highlighting a key limitation for future AI development.

Summary

Researchers at Goodfire.ai demonstrated that memorization and logical reasoning in large language models occupy distinct neural pathways. By surgically pruning low‑curvature weight components, they eliminated 97% of verbatim recall while preserving 95‑106% of reasoning performance on benchmarks such as BoolQ and OpenBookQA. The same pruning caused arithmetic ability to collapse to 66% of baseline, indicating that math operations share the memorization circuitry. The technique, validated on OLMo‑7B, OLMo‑1B, and vision transformers, offers a proof‑of‑concept for targeted removal of copyrighted or sensitive content without crippling core model functions.