AI Memory vs Reasoning Separate Neural Pathways

New research from AI startup Goodfire.ai suggests that AI language models, such as GPT-5, use distinctly separate neural pathways for memorisation and reasoning. This challenges the assumption that these functions are intertwined within the model's architecture.

The study found that by removing the "memorisation pathways," the AI models lost a significant 97% of their ability to directly recite training data. However, their "logical reasoning" capabilities remained almost entirely intact. This separation was observed by analysing the "loss landscape" of AI models, revealing that memorised facts create sharp spikes, while reasoning abilities maintain consistent curves.

Interestingly, arithmetic operations appear to be linked to memorisation pathways rather than logical reasoning. When the researchers removed memorisation circuits, mathematical performance notably declined, potentially explaining why AI language models often struggle with maths without external tools. This indicates that current models might treat simple arithmetic as memorised facts instead of computed operations. The research team tested their findings on various AI systems, including Allen Institute's OLMo-2 family and custom Vision Transformers, confirming the separation of memorisation and reasoning across different architectures. The K-FAC technique used in the study also outperformed existing memorisation removal methods.

While the research offers promising insights into AI functionality, the researchers acknowledge the limitations of their technique, noting that removed memories might return with further training and that the complete elimination of sensitive information cannot be guaranteed. Further, why some capabilities, like maths, are so easily disrupted by memory removal remains unclear. Nevertheless, these early findings open new avenues for AI research and development, potentially leading to methods for removing unwanted content from AI models without impairing their core functionalities.

Fuente Original: https://arstechnica.com/ai/2025/11/study-finds-ai-models-store-memories-and-logic-in-different-neural-regions/

Artículos relacionados de LaRebelión:

• Metas Smart Glasses Revolution Unveiling Hypernova with Display Neural Wristband - A First Look

Artículo generado mediante LaRebelionBOT