analiza aiqbench
This paper introduces "feature recall," a mechanism in deep learning where linear projections retrieve stored information scaled by input activations, extending beyond simple factual memory to a broader operational paradigm. By contrasting this with feature combination across various architectures, the work offers philosophers and researchers a new conceptual framework for mechanistic interpretability. It highlights specific empirical directions for understanding how models internally access and scale stored representations during inference.
Deep Learning Models Also Recall Features
arXiv:2608.20970v1 Announce Type: new Abstract: Recent work in mechanistic interpretability has studied how large language models recall facts stored in their weights. This paper argues that factual recall points to something broader: a general kind of operation in deep learning models, which I call feature recall. The core observation is that a linear projection can be read as retrieving stored information scaled by input activations. I define feature recall, show it applies across architectures, and contrast it with the established paradigm of feature combination. I also consider how cases of feature recall might be mechanistically identified. The account gives philosophers a new conceptual tool for understanding deep learning, and points to empirical directions for mechanistic interpretability research.
Abstract:Recent work in mechanistic interpretability has studied how large language models recall facts stored in their weights. This paper argues that factual recall points to something broader: a general kind of operation in deep learning models, which I call feature recall. The core observation is that a linear projection can be read as retrieving stored information scaled by input activations. I define feature recall, show it applies across architectures, and contrast it with the established paradigm of feature combination. I also consider how cases of feature recall might be mechanistically identified. The account gives philosophers a new conceptual tool for understanding deep learning, and points to empirical directions for mechanistic interpretability research.
Artificial Intelligence (cs.AI)
arXiv:2608.20970 [cs.AI]
(or arXiv:2608.20970v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2608.20970 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 206 cuvinte
Subiecte conexe
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · 8 oct. 2026
Hybrid Quantum Networks Lift Fraud Detection Accuracy to 85%
Quantum Zeitgeist · 4 aug. 2026
Game Design Creativity evals
LessWrong · 12 oct. 2026
Handling Imbalanced Classification: What Works Better Than SMOTE
Analytics Vidhya · 12 iul. 2026