Have you ever wondered how AI models manage to recall specific facts while also solving complex problems? It turns out, they’re not doing it the same way. Recent groundbreaking research from Goodfire.ai has revealed that AI neural networks handle memorization and reasoning through entirely separate pathways—a discovery that could reshape how we build and refine these systems. But here’s where it gets controversial: arithmetic, which we’d assume relies on logic, actually shares pathways with memorization, not reasoning. Could this be why AI models struggle with math? Let’s dive in.
The Divide Between Memorization and Reasoning
When engineers train AI language models like GPT-5, two key abilities emerge: memorization (reciting exact text they’ve encountered before) and reasoning (solving new problems using general principles). Goodfire.ai’s research, detailed in a preprint paper (https://arxiv.org/abs/2510.24256), provides the first clear evidence that these functions operate through distinct neural pathways. The separation is strikingly clean: when researchers removed the memorization pathways, models lost 97% of their ability to recite training data verbatim but retained nearly all their logical reasoning skills.
For instance, in the Allen Institute’s OLMo-7B language model (https://allenai.org/olmo), the bottom 50% of weight components showed 23% higher activation on memorized data, while the top 10% activated 26% more on general, non-memorized text. This clear split allowed researchers to surgically remove memorization without harming other capabilities.
Arithmetic’s Surprising Connection to Memorization
Here’s the part most people miss: arithmetic operations appear to rely on memorization pathways, not reasoning. When memorization circuits were removed, mathematical performance dropped to 66%, while logical tasks remained largely unaffected. This suggests AI models treat arithmetic more like recalling facts from a memorized table than performing logical computations. It’s akin to a student who’s memorized times tables but doesn’t understand multiplication itself. This finding could explain why AI models often falter at math without external tools (https://www.arsturn.com/blog/why-your-llm-is-bad-at-math-and-how-to-fix-it-with-a-clip-on-symbolic-brain).
What Counts as ‘Reasoning’ in AI?
It’s important to note that ‘reasoning’ in AI isn’t the same as human reasoning. The logical tasks that survived memory removal—like evaluating true/false statements or applying if-then rules—are essentially pattern matching, not deep mathematical reasoning. This contrasts with the kind of problem-solving required for proofs or novel solutions, which AI models still struggle with (https://arstechnica.com/ai/2025/04/new-study-shows-why-simulated-reasoning-ai-models-dont-yet-live-up-to-their-billing/).
The ‘Loss Landscape’: Mapping AI’s Inner Workings
To understand how researchers distinguished these pathways, consider the concept of the loss landscape—a way to visualize how an AI model’s predictions improve or worsen as its internal settings (weights) are adjusted. Imagine tuning a machine with millions of dials, where the ‘loss’ measures errors. High loss means many mistakes; low loss means few. The landscape maps the error rate for every possible combination of settings.
During training, AI models ‘roll downhill’ in this landscape, adjusting weights to minimize errors. Researchers analyzed the ‘curvature’ of these landscapes, measuring how sensitive performance is to small changes in weights. Sharp peaks and valleys indicate high curvature (tiny changes have big effects), while flat plains represent low curvature (changes have minimal impact).
Using a technique called K-FAC (https://arxiv.org/abs/1503.05671), they found that memorized facts create sharp spikes in the landscape, but these spikes point in different directions, averaging to a flat profile. In contrast, reasoning abilities maintain consistent moderate curves, like rolling hills that retain their shape regardless of approach.
Testing Across Systems and Tasks
The researchers tested their findings across multiple AI systems, including Allen Institute’s OLMo-2 models and custom Vision Transformers (ViT-Base models, https://huggingface.co/google/vit-base-patch16-224). They also compared their method to existing memorization removal techniques like BalancedSubnet (https://mansisak.com/memorization/).
When low-curvature weight components were removed, memorized content recall dropped to 3.4%, while logical reasoning tasks maintained 95-106% of baseline performance. These tasks included Boolean expression evaluation, logical deduction puzzles, and benchmarks like BoolQ (https://arxiv.org/abs/1905.10044) and Winogrande (https://github.com/allenai/winogrande).
Interestingly, mathematical operations and fact retrieval shared pathways with memorization, dropping to 66-86% performance after editing. Arithmetic proved especially fragile, even when models generated identical reasoning chains. Open-book question answering, which relies on provided context, remained nearly unaffected.
The Limits of Memory Removal
While promising, this technique isn’t perfect. Removed memories might return with further training, as current unlearning methods only suppress information rather than erase it (https://arxiv.org/pdf/2506.06278). Additionally, it’s unclear why math breaks so easily when memorization is removed—is it truly memorized, or does it just share neural circuits with memorization? Some complex reasoning patterns might also be misidentified as memorization.
Looking Ahead: Ethical and Practical Implications
If refined, this technique could allow AI companies to remove copyrighted content, private information, or harmful text from models without sacrificing functionality. However, neural networks store information in distributed, poorly understood ways, so complete elimination of sensitive data remains uncertain.
But here’s the thought-provoking question: If AI models rely on memorization for tasks like arithmetic, does this challenge our assumptions about their ‘intelligence’? And how can we redesign models to prioritize true reasoning over rote recall? Let us know your thoughts in the comments—this is a conversation worth having.
