Conceptlinear representation hypothesis
Linear Representation Hypothesis
A theoretical hypothesis about the internal representations of large models, holding that high-level semantic features are encoded as linear directions, with conceptual differences corresponding to vector differences and concept strength corresponding to the magnitude of projection along that direction.
Timeline (last 90 days)
Oct 23
The linear representation hypothesis mainly describes the separability of features rather than the compositional logic among features; how mappings across conceptual domains are realized through the coordination of multiple directions remains a part not yet covered by this hypothesis.
Unverified50%