The Manifold Turn in Mechanistic Interpretability- A Review of Geo-metric Approaches to Understanding Neural Network Internals
This paper mainly describes the turn of view from songle dimentional representation to manifold with a submanifold representation.
In linear prepresentation steering is really effective. But in manifold how can steering be employed ?will it be effective ? paper claims that linear steering isnt much effective Arditi et al., 2024) — appears to vindicate the strict linear picture.
non MI payoff payoff is the work of Kendiukhov (2026), who mechanistically inter-preted the single-cell foundation model scGPT and identified an internal hematopoietic manifold with developmental branching structure that closely tracks the known biology of blood-cell differ-entiation.
Limitations
© 2026 bsybin