Steering Geometry: Validating Human Value Geometry in LLM Steering Space
Some steering methods appear to preserve the expected map of human values, while others only match behavior.
The paper tests whether LLM steering vectors reflect Schwartz’s theory of human values, using a 26K-sample benchmark across 20 values. Distribution-driven methods showed value geometries aligned with the theory, with Spearman correlation up to 0.51. Behavior-centric methods steered outputs comparably but did not recover the same structure. The authors also report that scale improves geometric fidelity, instruction tuning weakens it, and better geometry transfers more consistently across compatible and opposing values. HF Daily Papers' note
The paper tests whether LLM steering vectors reflect Schwartz’s theory of human values, using a 26K-sample benchmark across 20 values. Distribution-driven methods showed value geometries aligned with the theory, with Spearman correlation up to 0.51. Behavior-centric methods steered outputs comparably but did not recover the same structure. The authors also report that scale improves geometric fidelity, instruction tuning weakens it, and better geometry transfers more consistently across compatible and opposing values. HF Daily Papers' note
score 4