Steering Geometry: Validating Human Value Geometry in LLM Steering Space

Steering Geometry. Validating Human Value Geometry in LLM Steering Space. By Mohammad Mahdi Abootorabi and seven others, from the University of British Columbia, the Vector Institute for AI, and one other institution. Published on the fifth of September. Its code repository is worked out: about two hundred source files, with a readme.

Activation steering has become a lightweight, inference-time alternative to fine-tuning methods such as RLHF and DPO for controlling how large language models behave, and those models are increasingly deployed where alignment matters. But steering is usually validated one isolated behaviour at a time. That leaves open whether steering vectors encode coherent semantic structure, or merely exploit shortcuts specific to each behaviour. The question here is whether the latent geometry of steering vectors reflects the structure that theory specifies for human values and morality. The answer depends on the method. Distribution-driven methods recover that structure; behaviour-centric methods largely do not.

The primary framework is Schwartz's Theory of Basic Human Values, used for its fine grain. The authors build a benchmark of twenty-six thousand samples covering twenty human values. They analyse distribution-driven methods such as CAA, SphericalSteer and ODESteer, and behaviour-centric approaches such as COLD-Steer and BiPO, across diverse model families and sizes.

Distribution-driven methods recover value topologies aligned with the theory's predictions, with a Spearman correlation of up to zero point five one and a p-value below ten to the minus thirteen. Behaviour-centric methods steer comparably well, but their vectors show little correlation with the expected value geometry. Geometric fidelity improves with model scale and drops after instruction tuning. And better geometric alignment brings more human- consistent transfer across values: steering one value correctly lifts compatible values and suppresses opposing ones.

The claims are relative to one theoretical map of values, used as the primary framework, and to the methods and model families tested. Correlations reach up to zero point five one, which is alignment with the theory, not identity with it. And steering performance alone does not reveal geometry: two methods can steer equally well while only one respects the structure.

That last point is the practical one. If steering success is the only test, a shortcut and a coherent representation look the same. Checking geometry against a theory of values separates them, and the geometry turns out to matter for how steering one value spills over into the rest. Code and data are available.