Partial least squares (PLS) subspace probing

methods / Direction Extraction / Partial least squares (PLS) subspace probing

Partial least squares (PLS) subspace probing

Techniqueintermediate

Fits a multi-component partial least squares regression from activations to a target numeric variable, identifying a low-dimensional linear subspace (not just a single direction) that predicts the variable, then uses the leading component as an intervention direction.

Used in (1 observation)

PLS-found numeric-comparison subspaces causally control LLM answersmeasured

structure: Linear Subspace · models: Llama-3-8B-Instruct, Mistral-7B-Instruct, Qwen2.5-7B-Instruct · paper: The Geometry of Numerical Reasoning: Language Models Compare Numeric Properties in Linear Subspaces