Spatial relation words in LLaMA and Qwen models form antipodal, roughly orthogonal directions in a low-dimensional subspace, and steering along these directions changes the model's output.
World-model interpretability is all we need.AI Alignment Forum, January 2023
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.AI 1years
2025 1verdicts
REJECT 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Linear Spatial World Models Emerge in Large Language Models
Spatial relation words in LLaMA and Qwen models form antipodal, roughly orthogonal directions in a low-dimensional subspace, and steering along these directions changes the model's output.