Sheaf-Theoretic Decomposition for Transformer Interpretability
October 11, 2026
A new method uses cellular sheaves over paraphrase graphs to identify a content subspace in frozen models. This approach outperforms the LEACE concept eraser by 17.1 percentage points on CounterFact fact retrieval across eight architectures.
HOW THIS AFFECTS YOU
●
researcherYou can use this topological approach to better identify and manipulate semantic content within transformer activations.