Steering Geometry: Validating Human Value Geometry in LLM Steering Space

AI summary

Overview

Research area: Mechanistic interpretability and activation steering for large language models, evaluated against psychological theories of human values (Schwartz's Theory of Basic Human Values and Moral Foundations Theory).

Technical level: Advanced. The paper assumes familiarity with residual-stream activations, steering vectors, contrastive activation addition, sparse autoencoders, and correlation-based geometry metrics.

Scope: A systematic study of whether the latent geometry of LLM steering vectors reflects the theory-specified structure of human values, comparing distribution-driven versus behavior-centric steering methods across model families, scales, and tuning regimes.

What This Paper Is About

Activation steering lets researchers control LLM behavior at inference time by adding directions to internal activations, but existing work validates these vectors only on the single behavior they target. The authors ask whether steering vectors encode coherent semantic structure — specifically, whether their pairwise similarities match the circular arrangement of values predicted by Schwartz's Theory of Basic Human Values — or whether they merely exploit behavior-specific shortcuts. To answer this, they build a contrastive benchmark of roughly 26K samples covering 20 human values and use it to measure the geometric fidelity of steering vectors produced by seven different methods.

Key Contributions

1. First study of steering-vector geometry against human value theory. The authors compare distribution-driven and behavior-centric steering methods against the Schwartz circumplex, finding that distribution-driven methods recover the expected value topology (Spearman ρ up to 0.51, p < 10⁻¹³) while behavior-centric methods show near-zero correlation despite achieving comparable steering performance. The same paradigm-level separation appears under Moral Foundations Theory.

2. Evidence that model scale and tuning regime affect geometric fidelity. Larger and more recent model families show stronger geometric alignment, whereas instruction-tuned models show weaker value geometry, which the authors describe as "value geometry drift" associated with post-training.

3. A new benchmark of roughly 26K contrastive quadruples spanning the Schwartz taxonomy, built from 911 samples extracted from ValueBench and around 25.5K samples from the Touché data, using the 20-value taxonomy of Kiesel et al. (2022) that extends Schwartz's refined 19 values with Universalism: Objectivity. A second balanced benchmark of 1.2K question-contrastive samples covers the six Moral Foundations Theory foundations using the Moral Foundations Reddit Corpus.

4. A cross-value transfer analysis showing geometry predicts behavior. Methods that better reflect human value geometry also transfer more consistently across the value space: steering toward one value correctly gains accuracy on compatible values and suppresses opposing ones, while behavior-centric methods cluster near zero on transfer fidelity.

Main Findings

• Distribution-driven methods match theory; behavior-centric methods do not. On the Qwen 3.5-9B backbone (Base), SAS achieves the best geometry alignment with Theory Rank Correlation ρ_T = 0.5069 (p = 8.5e-14), Theory Linear Correlation r_T = 0.5061 (p = 9.4e-14), Hierarchical Structure Correlation ρ_H = 0.3910 (p = 2.4e-8), and Polarity Separation Score Δ_pol = 0.3951. CAA follows at ρ_T = 0.4606 (p = 2.3e-11), SphericalSteer at 0.3962 (p = 1.5e-8), and ODESteer at 0.2730 (p = 1.4e-4). By contrast, OPT reaches only ρ_T = 0.1138 (p = 1.2e-1), COLD-Steer (FD) 0.0265 (p = 7.2e-1), and BiPO 0.1188 (p = 1.0e-1) — none statistically significant.

• Raw activations sit between the two paradigms. The unsteered LLM raw activation space on Qwen 3.5-9B-Base scores ρ_T = 0.2228 (p = 2.0e-3), r_T = 0.2087 (p = 3.9e-3), ρ_H = 0.2115 (p = 3.4e-3), and Δ_pol = 0.0009. Distribution-driven methods "sharpen" this partial structure; behavior-centric methods weaken it.

• Instruction tuning degrades value geometry across all methods. On Qwen 3.5-9B-Instruct, SAS drops to ρ_T = 0.3256, CAA to 0.2351, SphericalSteer to 0.2048, ODESteer to 0.2055, and the raw activation baseline to 0.1253. OPT falls to 0.0615, COLD-Steer to −0.0568, and BiPO to −0.0104.

• Geometric fidelity improves with model scale. Within the Qwen3.5 and Gemma-3 families, metrics improve monotonically with scaling for distribution-driven methods. Raw activations improve only modestly, and behavior-centric methods show no reliable scale trend.

• Newer model families align better, but geometric fidelity is independent of downstream performance. At the 7–9B range, ρ_T rises from Falcon-7B (0.20) to Mistral-7B (0.32), Llama-3.1-8B (0.37), and Qwen3.5-9B (0.46). Qwen3.5 outperforms Qwen2.5 across all scales, and Qwen models outperform Gemma models at comparable scales. Gemma-4-31B, despite being one of the strongest open models, achieves a lower ρ_T (0.38) than Qwen3.5-4B (0.41).

• The paradigm split generalizes beyond Schwartz. Under MFT, distribution-driven methods yield positive family separation on both backbones (Δ_MFT ≈ 0.04–0.10 versus ≈0.001 for raw-activation baselines), while behavior-centric methods remain near zero or negative. On Qwen3.5-9B / Llama3.1-8B respectively: CAA 0.0969/0.1038, SphericalSteer 0.0932/0.0909, SAS 0.0892/0.0923, ODESteer 0.0495/0.0378, raw activations 0.0006/0.0017, OPT 0.0107/0.0272, BiPO −0.0287/0.0227, and COLD-Steer −0.0064/−0.0280.

• Geometry predicts cross-value transfer better than raw accuracy does. Pooling all methods and both backbones, geometric fidelity (ρ_T) predicts transfer more strongly than raw accuracy gain on both TWTM and ρ_H^tr, with the gap widening on the rank-based hierarchical metric (Δρ = 0.06 on TWTM, Δρ = 0.18 on ρ_H^tr).

• Qualitative analysis confirms the pattern. In cosine-similarity heatmaps and 2D projections on Qwen3.5-9B-Base, SAS most closely reproduces the theoretical Schwartz matrix, raw activations preserve a partial but recognizable pattern, and OPT remains largely unstructured — described as "worse than the raw-activation baseline."

• Robustness checks preserve the conclusion. The paradigm-level separation holds when re-extracting vectors from a fresh disjoint subset of 200 examples per value, when fully paraphrasing questions and answers, and when restricting to a single-label subset. Absolute correlations decrease in the single-label setting, but the matched raw activation baseline decreases as well.

Methodology in Plain English

The study runs a four-stage pipeline.

Stage 1 — Build a contrastive dataset. Because most steering methods need contrastive pairs to extract value directions, the authors assemble (question, value, positive answer, negative answer) quadruples where the positive answer aligns with a target value and the negative answer is value-neutral. They combine 911 samples from ValueBench with roughly 25.5K samples from Touché, yielding almost 26K data points across 20 Schwartz values, sampled at 200 contrastive prompt pairs per value with a 90/10 train/test split.

Stage 2 — Extract value directions. For each steering method and each value, the authors define a single value direction vector as the expected shift the method induces in the residual-stream hidden state at layer ℓ, averaged over input prompts. This common formulation lets them compare methods with very different internal mechanisms. The vectors for all 20 values are collected into a per-method "value vector bank." The intervention layer is chosen by maximizing a combined criterion of Normalized L2 Separation and Linear Probe Accuracy — the layer where positive and negative activations are most linearly separable. Hyperparameters are finetuned per method on a validation split.

Stage 3 — Test geometry. The value vectors are mean-centered, normalized to unit length, and turned into an empirical cosine-similarity matrix. This is compared against a theoretical matrix where each entry is the cosine of the angular separation between values on the Schwartz circumplex (each step being 18 degrees). Four metrics quantify the match: Spearman rank correlation (ρ_T), Pearson correlation (r_T), Hierarchical Structure Correlation (ρ_H) using a four-level distance (1 = same sub-family, 2 = same higher-order group, 5 = unrelated, 10 = opposing groups), and Polarity Separation Score (Δ_pol), which contrasts within-family versus opposing-group similarity.

Stage 4 — Test cross-value transfer. For each ordered pair of distinct Schwartz values, the authors steer toward one value and evaluate accuracy on the other's held-out test set, producing a 20 × 20 matrix per method with 380 off-diagonal entries. Because raw transfer scores mix row effects (some source values steer strongly) and column effects (some targets are easy to improve), they apply two-way centering to isolate pair-specific signal. Two metrics then quantify alignment: Continuous Transfer Fidelity (TWTM), which weights residualized transfer entries by theoretical Schwartz affinity, and Hierarchical Transfer Fidelity (ρ_H^tr), a Spearman correlation against the negative hierarchical distance. A shape-only robustness metric, Bin-wise Monotonic Decay (BMD-ρ), is also reported.

Models and methods tested. CAA is evaluated across seven pre-trained model families: Llama3.1, Mistral, Falcon, Qwen2.5, Qwen3.5, Gemma-3, and Gemma-4. All steering methods are applied to Qwen3.5-9B in both pre-trained (Base) and instruction-tuned (Instruct) variants, and to Llama3.1-8B (Base), for controlled comparison. Accuracy gain is measured as the change in the model's selection rate of the value-aligned answer in a two-option multiple-choice setup relative to the unsteered baseline.

Why This Matters

Impact on research. The paper argues that evaluating steering vectors only on their target behavior is insufficient, because behavior-centric methods can hit the target while destroying the semantic structure that makes an intervention predictable. It offers a theory-grounded test of whether steering vectors encode meaning or exploit shortcuts, and it supplies a benchmark and metrics to make that test repeatable. It also reframes instruction tuning as something that can erode latent value structure even as it improves surface behavior.

Real-world applications:

• Alignment auditing. Regulators and auditors could use geometric fidelity as a diagnostic for whether a model's internal value representations are coherent, rather than relying only on behavioral test suites.
• Safer deployment of steerable assistants. Applications that steer models toward empathy, honesty, or safety at inference time would benefit from knowing whether steering one trait silently suppresses or amplifies unrelated ones.
• Culturally sensitive alignment. Because Schwartz's circumplex is described as cross-culturally validated, geometry-based evaluation could help assess whether a model's value representations generalize across cultural contexts rather than being tuned to a single population.
• Model selection and comparison. The finding that a stronger model (Gemma-4-31B, ρ_T = 0.38) can have weaker value geometry than a smaller one (Qwen3.5-4B, ρ_T = 0.41) gives practitioners a reason to evaluate representational structure independently of benchmark accuracy.

Industry relevance. Teams that ship steering-based behavioral control — rather than full fine-tuning — need interventions whose side effects are predictable. The result that better geometric alignment also yields more human-consistent cross-value transfer suggests that geometry metrics could serve as a cheap pre-deployment screen, and that choosing distribution-driven extraction (CAA, SAS, SphericalSteer, ODESteer) over behavior-centric optimization (OPT, BiPO, COLD-Steer) trades no accuracy for substantially more predictable behavior.

Future Directions

1. Extend beyond Schwartz and MFT. MFT specifies only the coarser Individualizing–Binding grouping and no finer inter-foundation structure, so it provides family-level validation rather than a full geometric test. Generalization to other value frameworks, and across cultures and languages, remains open; the authors note that cognitively grounded evaluations reveal substantial cross-lingual gaps.

2. Address the monosemanticity gap. Real arguments may express multiple values, so the extracted vectors should be read as target-conditioned aggregate directions rather than perfectly monosemantic representations. The single-label-prioritized sensitivity test preserves the paradigm-level conclusion but does not establish perfect monosemanticity.

3. Broaden cross-backbone comparisons. Most cross-backbone comparisons use only a few steering methods due to computational constraints, leaving open whether the observed trends hold consistently across other methods.

4. Test value-level versus behavior-level steering. The authors propose examining whether steering through values produces more predictable downstream effects than steering surface behaviors directly — for example, whether steering politeness introduces side effects such as suppressing honesty or amplifying conformity.

Target Audience

Researchers and engineers working on LLM alignment, mechanistic interpretability, and representation engineering — particularly those building or deploying activation-steering interventions. The paper is also relevant to AI safety and policy audiences interested in auditing tools for model behavior, and to computational social scientists using psychological frameworks such as Schwartz's Theory of Basic Human Values or Moral Foundations Theory to study model representations. The content is technical, requiring comfort with activation-space analysis and correlation-based evaluation, so it is best suited to readers with prior exposure to interpretability methods rather than newcomers.

Authors’ abstract

As large language models (LLMs) are increasingly deployed in alignment-sensitive contexts, activation steering has emerged as a lightweight, inference-time alternative to fine-tuning methods (e.g., RLHF, DPO) for behavioral control. However, existing work typically validates steering on isolated behaviors, leaving it unclear whether steering vectors encode coherent semantic structure or merely exploit behavior-specific shortcuts. We investigate whether the latent geometry of LLM steering vectors reflects theory-specified structure in human values and morality. Using Schwartz's Theory of Basic Human Values as our primary fine-grained framework, we introduce a 26K-sample benchmark covering 20 human values and analyze distribution-driven methods (e.g., CAA, SphericalSteer, ODESteer) and behavior-centric approaches (e.g., COLD-Steer, BiPO) across diverse model families and sizes. We find that distribution-driven methods recover human value topologies aligned with theoretical predictions (Spearman $ρ$ up to 0.51, $p &lt; 10^{-13}$). In contrast, behavior-centric methods achieve comparable steering performance but show little correlation with the expected value geometry. Geometric fidelity improves with model scale but drops after instruction tuning. Finally, better geometric alignment also leads to more human-consistent transfer across values: steering one value correctly lifts compatible values and suppresses opposing ones. Code and data are available at: https://github.com/DeepRCL/Steering_Geometry.