Reproduce “Steering Geometry: Validating Human Value Geometry in LLM Steering Space” on OpenResearch

About this paper

Mohammad Mahdi Abootorabi, Armin Saghafian, Ali Bazshoushtari, Hamid Rezaei, EunJeong Hwang, Vered Shwartz, Parvin Mousavi, Purang Abolmaesumi

The authors test whether activation-steering directions preserve Schwartz’s relationships among 20 values and whether their geometry tracks cross-value transfer. They compare distribution-driven and behavior-centric methods using direction similarities and value-aligned answer selection. On Qwen3.5-9B-Base, SAS and OPT achieve similar target-accuracy gains, +10.03 and +9.57 points, but theory correlations differ, 0.5069 and 0.1138. Results concern tested methods, backbones, and Schwartz’s structure.

Key insights

• Distribution-driven steering captures value relationships more consistently than methods optimized directly for target-answer probabilities, even when both achieve similar target behavior gains.
• Geometric fidelity is associated with predictable transfer: steering toward one value tends to benefit neighboring values and suppress opposing ones when the extracted directions follow the circumplex.
• The structure is model-dependent: alignment generally increases with scale within families but varies across architectures, and is lower in instruction-tuned models than in corresponding base models.

Summary by alphaXiv.