Adding or subtracting a direction in the residual stream during generation to change behavior, and reading off what changes. The main causal test for whether a representation does anything.
Used in
- Han Chalmers Izmailov 2026 - Functional welfare axis
- Gilg et al 2026 - Probing persona-dependent preferences
- Tagliabue Dung Berg 2026 - The pain axis
- Sofroniew et al 2026 - Emotion concepts
- Lu et al 2026 - The Assistant Axis
- Chen et al 2025 - Persona vectors
- Sun et al 2026 - Valence-arousal subspace
- Bianco Shiller 2026 - Mechanistic tracing of pain-pleasure decisions
- Gurnee Sofroniew et al 2026 - A global workspace in language models
- Berg de Lucena Rosenblatt 2025 - Self-referential processing
- Black Bloom 2026 - Machinic psychopharmacology
- Kirk et al 2025 - Relationship-seeking AI