← Back to archive

How LLMs Separate Memorization from Reasoning

14 November 2025 · 40 views · 0 likes · 0 reposts · 120 blog views

AI chatgpt learning neuroscience research

An article breaks down how models actually separate memorization from reasoning. It shows that these two abilities run through totally different neural pathways inside LLMs.

If you’re curious how these things think, there are a few parts:

1️⃣ Memorization and reasoning live in different parts of the model

You can remove the memorization circuits and the model will forget almost everything it memorize while keeping its problem-solving skills nearly untouched. 97% of memorized content gone but logical reasoning still intact ‼️

2️⃣ for LLMs math isn’t reasoning - it’s memorization

If you remove the memorization pathways, math performance drops to ~66%. Models don’t calculate the way we think - they recall math facts from a mental lookup table. So 2 + 2 = 4 behaves more like a quote from a book than an actual computation.

3️⃣ Curvature of the "loss landscape" to map these pathways

High-curvature components = shared reasoning mechanics. Low-curvature components = sharp, isolated spikes - perfect fingerprints for memorized facts.

This finally gives researchers a way to find and separate the two.

4️⃣ Selective "unlearning" might be possible

Eventually you'll be able to remove copyrighted text/private data/harmful memorized content from a model without breaking everything else. Not there yet, but could be a step in that direction.

5️⃣ Not everything fits neatly

  • Some "forgotten" info can come back with retraining
  • Math still doesn’t make sense mechanistically
  • Rare facts disappear more than common facts
  • Some "reasoning" tasks may actually be disguised memorization

The line between reasoning and memory isn’t as simple as it looks.

How LLMs Separate Memorization from Reasoning

Originally posted on X

PS: I'm pretty active on 𝕏 if you'd like to follow along

Related posts

Archive

2026

2025

2024

2018