Spherical Steering: Geometry-Aware Activation Rotation for Language Models¶

Conference: ICML 2026
arXiv: 2602.08169
Code: https://github.com/chili-lab/Spherical-Steering (Available)
Area: Interpretability / Activation Editing / Inference-time Intervention / LLM Alignment
Keywords: Activation Steering, Hyperspherical Geometry, Slerp Geodesic, vMF Confidence Gating, Norm Preservation

TL;DR¶

This paper proposes Spherical Steering: rotating activation vectors along geodesics on the unit hypersphere of LLM hidden states toward a "truthfulness direction" estimated from contrastive samples. Unlike traditional additive activation steering, this method preserves activation norms while significantly improving multiple-choice accuracy on benchmarks like TruthfulQA, COPA, and StoryCloze (+10% range) without degrading the quality of open-ended generation.

Background & Motivation¶

Background: To control LLM behavior without retraining, the mainstream approach is activation steering—estimating a "steering vector" \(\mu\) from a batch of (positive, negative) contrastive samples and adding it directly to token activations at certain layers: \(h' = h + \lambda \mu\). Representative methods include CAA and ITI.

Limitations of Prior Work: This additive operation suffers from severe scale sensitivity. If \(\lambda\) is too small, there is no effect; if \(\lambda\) is large, the hidden state norm \(\|h\|\) is significantly distorted—the equation \(\|h'\|^2 = \|h\|^2 + 2\lambda\mu^\top h + \lambda^2\) shows that norm changes depend on both \(\lambda\) and the alignment between \(\mu\) and \(h\), making them uncontrollable. Consequently, while multiple-choice accuracy may increase, open-ended generation quality (TRUE×INFO) often drops, leading to "over-conservatism" or even representation collapse.

Key Challenge: Modern LLMs commonly normalize activation magnitudes using RMSNorm/LayerNorm, implying that direction is the primary degree of freedom carrying semantics. Additive steering perturbs magnitudes freely, conflicting with the geometric priors of the architecture.

Goal: Design a geometry-consistent inference-time intervention primitive that is training-free like addition but strictly preserves \(\|h\|\), avoiding damage to the geometric priors of normalization layers.

Key Insight: The authors made a critical empirical observation (Fig. 3)—on TruthfulQA, the \(\ell_2\) norm curves of last-token activations for "correct" and "incorrect" answers almost overlap across all 32 layers (difference <1%), but their directions differ significantly. This indicates that truthfulness signals are encoded in direction rather than magnitude.

Core Idea: Normalize activations to the unit hypersphere \(\mathbb{S}^{d-1}\), rotate them toward a target direction \(\mu_T\) along a geodesic (great circle) using Slerp, and finally scale back to the original norm. This norm-preserving rotational intervention essentially replaces "addition in \(\mathbb{R}^d\)" with "rotation on \(\mathbb{S}^{d-1}\)".

Method¶

Overall Architecture¶

The method consists of three training-free steps: 1. Offline Prototype Construction: Run the model on a contrastive dataset \(\mathcal{D}=\{(x_i, y_i^+, y_i^-)\}\), extract last-token activations from layer \(l\), compute the mean difference, and normalize it to obtain the "truthfulness axis" \(\mu_T^{(l)}\) for that layer. 2. Inference-time Rotation: For each intervened layer \(l\) and each decoded token \(j\), normalize the current activation \(h_j^{(l)}\) to the sphere, rotate it toward \(\mu_T^{(l)}\) via Slerp with a gating-determined step size \(t_j^{(l)} \in [0,1]\), and restore the original norm. 3. vMF Confidence Gating: Use the exponential form of the von Mises–Fisher distribution to perform a two-class softmax on the similarity between the current direction and \((\mu_T, \mu_H = -\mu_T)\), obtaining a "hallucination bias" confidence \(\delta\). This is mapped to the intervention strength \(t\) via threshold \(\beta\) and scaling \(\alpha\)—applying strong rotation only when the model "appears to hallucinate."

Key Designs¶

Hyperspherical Prototype and Contrastive Mean Direction:
- Function: Extract the "truthfulness direction unit vector" \(\mu^{(l)}\) for a layer from contrastive samples at once.
- Mechanism: Feed concatenated sequences \(x_i \| y_i^\pm\) for each \((x_i, y_i^+, y_i^-)\) into the model and take the last-token representation \(z_i^{(l)\pm}\) of layer \(l\). Compute the difference between positive and negative means \(\Delta^{(l)} = m_+^{(l)} - m_-^{(l)}\) and normalize it: \(\mu^{(l)} = \Delta^{(l)}/\|\Delta^{(l)}\|\). This step is offline, keeps model weights frozen, and is performed only once.
- Design Motivation: The mean difference automatically suppresses shared context between positive and negative samples, highlighting the discriminative component of "truth vs. falsehood." Normalization is required because subsequent operations occur on \(\mathbb{S}^{d-1}\), necessitating a pure "direction" rather than a scaled offset. This is lighter than ITI's per-head probes and more geometrically consistent than CAA's additive approach.
Geodesic Rotation = Slerp + Norm Restoration:
- Function: Rotate \(h^{(l)}\) along the shortest spherical path toward \(\mu_T\) by a ratio \(t\), then restore the original magnitude.
- Mechanism: Compute the angle \(\theta = \arccos(\mu_T^\top \hat h^{(l)})\), then apply Shoemake’s (1985) spherical linear interpolation: \(\hat h^{(l)\prime} = \frac{\sin((1-t)\theta)}{\sin\theta}\hat h^{(l)} + \frac{\sin(t\theta)}{\sin\theta}\mu_T\). Finally, set \(h^{(l)\prime} = \|h^{(l)}\|\hat h^{(l)\prime}\). When \(t=0\), there is no change; when \(t=1\), it aligns perfectly with \(\mu_T\). Degenerate cases (\(\theta=0/\pi\)) are handled separately.
- Design Motivation: Slerp provides the path of minimal angular change for a given step \(t\), meaning "maximum semantic alignment with minimal directional perturbation." Meanwhile, \(\|h^{(l)\prime}\| \equiv \|h^{(l)}\|\) holds strictly, bypassing the issue of uncontrolled \(\|h\|\) in additive steering and conforming to the architectural prior that direction carries information after RMSNorm. Unlike Angular Steering which rotates in a fixed 2D plane, this method performs hyperspherical rotation in the original \(d\)-dimensional space without relying on PCA projections.
Input-Adaptive Step Size via vMF Confidence Gating:
- Function: Vary \(t\) based on the token—no intervention if the model is already in the "truthful" hemisphere, and stronger intervention as it leans toward the "hallucination" hemisphere.
- Mechanism: Use the exponential term of the vMF density \(f(u;m,\kappa)\propto\exp(\kappa m^\top u)\) as a prototype score. Perform a two-class softmax over \((\mu_T, \mu_H)\) to get \(p_T, p_H\), and define \(\delta = p_H - p_T \in [-1,1]\). Apply threshold \(\beta\) and scaling \(\alpha\): \(t = \mathrm{clip}(\alpha \cdot \frac{\delta-\beta}{1-\beta}, 0, 1)\), where \(t=0\) if \(\delta \le \beta\).
- Design Motivation: Compared to a uniform \(t\) for all tokens, gating provides two benefits (Ablation Fig. 5): higher MC accuracy peaks with wider intervals, and stable TRUE×INFO even at high intensities (stable at \(\alpha=1.0\)), whereas ungated performance collapses at \(\alpha > 0.6\). \(\kappa\) controls the steepness of the confidence curve. This design follows the principle of "applying intervention only where a hallucination is likely."

Inference Flow¶

For each decoding step and each selected layer \(l\) in \(\mathcal{L}=\{l_1,\dots,l_K\}\): Extract \(h_j^{(l)}\) → Normalize → Compute \(s_T, s_H\) → vMF gate computes \(t_j^{(l)}\) → If \(t>0\), perform Slerp rotation and restore norm; otherwise, pass through. The complexity involves a few dot products and sin/cos operations, which are negligible compared to the forward pass.

Key Experimental Results¶

Main Results¶

On TruthfulQA (LLaMA-3.1-8B-Instruct), Spherical Steering achieves the best performance across multiple-choice metrics (MC1/MC2/MC3) and open-ended generation (TRUE×INFO) simultaneously—whereas additive baselines like ITI/CAA improve MC at the cost of TRUE×INFO, showing a typical trade-off.

Model	Method	MC1	MC2	MC3	TRUE×INFO
LLaMA-3.1-8B-Instruct	Baseline	34.15	53.32	27.02	48.24
LLaMA-3.1-8B-Instruct	ITI	37.70	58.09	30.12	40.31 ↓
LLaMA-3.1-8B-Instruct	CAA	35.99	56.26	29.36	49.66
LLaMA-3.1-8B-Instruct	SADI-HEAD	38.53	56.03	30.57	51.18
LLaMA-3.1-8B-Instruct	Spherical (Ours)	49.95	68.51	41.05	54.63
Qwen-2.5-7B-Instruct	Baseline	35.87	54.95	26.62	74.40
Qwen-2.5-7B-Instruct	ITI	40.15	58.93	30.26	67.82 ↓
Qwen-2.5-7B-Instruct	Spherical (Ours)	48.71	66.90	39.16	77.84

Zero-shot evaluation across 6 multi-choice benchmarks (LLaMA-3.1-8B-Instruct):

Method	TruthfulQA	COPA	StoryCloze	MMLU	Wino.	BoolQ	Avg.
Baseline	34.15	83.00	74.72	60.60	50.81	80.12	63.90
ITI	37.70	83.00	75.12	60.90	51.85	81.53	65.02
CAA	35.99	84.00	79.02	60.70	51.93	82.42	65.68
SADI-HEAD	38.53	84.00	75.72	60.66	51.85	80.20	65.16
Spherical (Ours)	49.95	95.00	89.08	62.05	52.72	82.94	71.96

Average absolute Gain of +6.28%, with > +10% on COPA/StoryCloze.

Ablation Study¶

Configuration	MC1 (TruthfulQA, LLaMA)	TRUE×INFO	Description
K=1 layer	45.41	52.16	Single layer rotation: MC already near peak
K=2 layers	47.62	73.93	Additional layers mainly recover generation quality (INFO 62.9→90.3)
K=3 layers	47.13	74.43	Best overall balanced point
K=4 layers	41.37	70.62	Excessive intervention hurts MC
K=5 layers	41.37	70.09	Same as above
Ungated rotation (α=1.0)	—	Sharp decrease	Generation quality collapses at high α
vMF gated (α=1.0)	—	Remains stable	Gating significantly extends the usable range of α

Key Findings¶

Geometric Insight: Fig. 3 shows that the activation norms for truthful vs. hallucinated instances overlap almost perfectly across all layers (<1% difference), proving that truthfulness signals reside in direction rather than magnitude, empirically validating the norm-preserving design.
Collapse-Efficiency Advantage: Fig. 4 shows that for the same decrease in effective rank (Δrank≈50), rotation gains 8–10% more MC accuracy than addition. While addition sees TRUE×INFO collapse after a slight rank decrease, rotation sustains quality gains across a wide range of rank drops.
Asymmetric Effects of Multi-layer Intervention: Moving from K=1 to K=3 keeps MC almost constant (+2.2%) but jumps INFO from 62.9% to 92.7%. The authors suggest middle layers govern semantic discrimination (MC signal), while later layers govern token-level generation dynamics (INFO signal).
Orthogonality to 5-shot ICL: When combined with ICL, ITI drops TRUE×INFO from 38.9 to 37.3, while Spherical simultaneously pushes MC1 to 52.4% and TRUE×INFO to 42.8%, indicating that geometric intervention and prompt engineering operate via independent mechanisms.
High Sample Efficiency: Using only 25 contrastive samples increases MC1 on LLaMA from 36.3% to 51.5% (±2.2), with variance shrinking rapidly as samples increase.

Highlights & Insights¶

Reframing "Addition in \(\mathbb{R}^d\)" as "Rotation on \(\mathbb{S}^{d-1}\)" is a natural yet overlooked perspective: Since architectures like RMSNorm already stabilize norms, the remaining degrees of freedom for free perturbation are purely directional. This work fully implements this observation as an intervention primitive.
Slerp appears in LLM steering for the first time in a closed-form, training-free manner: Unlike methods like HPR that learn a Householder reflection, Spherical achieves both "geometric consistency" and "zero training."
vMF Gating is a lightweight plugin transferable to any steering method: It essentially uses the interpretable confidence of direction to dynamically adjust intensity, potentially applicable to CAA, ITI, or SAE-based interventions to decouple norm and direction control.
Strong "Pareto Improvement" arguments: Fig. 1(a) plots MC accuracy against TRUE×INFO, showing all baselines stuck on a trade-off curve while the proposed method jumps to the upper-right—a compelling demonstration of breaking the trade-off.
The concept of collapse-efficiency is methodologically valuable: Instead of just end-point metrics, the introduction of "performance gain per unit of rank decrease" provides a comparative geometric efficiency metric for future intervention research.

Limitations & Future Work¶

Prototypes rely on binary contrastive data: Currently supports binary concepts (truthful/hallucinated, safe/unsafe). Extending this to "multi-class fine-grained concepts" (e.g., multiple emotions or styles) would require multi-prototype or multi-axis geometry, which is not discussed.
Strong assumption of uni-axial \(\mu_T\) and its antipode \(\mu_H = -\mu_T\): In reality, "truth" may not be perfectly antipodal to "hallucination," potentially causing failure in tasks with mixed correct/incorrect answers.
Layer selection still relies on grid search: While the method selects layers \(\mathcal{L}=\{l_1,\dots,l_K\}\), the optimal combination is determined empirically (e.g., layer 24 for LLaMA), lacking a principled selection criterion.
Validated only on 7–8B Instruct models: Robustness of the hyperspherical assumption on base models, larger scales (30B+), or MoE architectures is unknown.
Gating hyper-parameters: \(\kappa, \alpha, \beta\) collectively determine the gating shape, representing a non-trivial tuning space. Automatically estimating \(\kappa\) from contrastive samples (vMF MLE) would be more efficient.
Improvement Directions: (i) Expanding the single axis \(\mu_T\) into low-rank multi-axial geometry for composite concept steering; (ii) Using SAE features as prototype directions; (iii) Replacing "geodesics" with Riemannian gradient flow for multi-step iterative rotation.

vs. CAA (Rimsky et al., 2024): CAA uses layer-wise addition \(h + \lambda\mu\). This work replaces it with Slerp rotation to preserve the norm. On LLaMA, CAA achieves MC1=35.99 / TRUE×INFO=49.66, while this work achieves 49.95 / 54.63.
vs. ITI (Li et al., 2023): ITI selects "truthful heads" via per-head linear probes and applies small additions. This work uses whole-layer directional rotation. While ITI's TRUE×INFO drops to 40.31 on LLaMA, this work increases it to 54.63.
vs. Angular Steering (Vu & Nguyen, 2025): Also an angular intervention, but it projects activations into a fixed 2D plane first, relying on low-dimensional kernels. This work operates directly on the raw \(d\)-dimensional spherical geodesic without PCA.
vs. HPR (Pham & Nguyen, 2024): HPR uses Householder reflections and a trained angle prediction network. This work is closed-form and training-free, sacrificing per-input angle learning flexibility for lightweight vMF-based adaptation.
vs. ReFT / LoFiT (Wu et al., 2024; Yin et al., 2024): These involve representation fine-tuning with lightweight modules. This work takes the "structured intervention" idea to a training-free extreme using pure geometric priors.
Insight: This "Sphere + Geodesic + Confidence Gate" combination can be transferred to any scenario where semantics are directionally encoded—image tokens in VLMs, noise embeddings in diffusion models, or graph representations. Any layer following LayerNorm/RMSNorm where editing is required should consider "Addition vs. Rotation" for geometric consistency.

Rating¶

Novelty: ⭐⭐⭐⭐ Replacing addition with rotation is a single-point idea, but the complete combination of hyperspherical geometry, Slerp, and vMF gating with rigorous geometric proof makes it a clean and effective innovation.
Experimental Thoroughness: ⭐⭐⭐⭐ Covers 6 MC benchmarks, open-ended generation, collapse-efficiency, and ablations for layers/gating/ICL/sample size. Lacks validation on larger models.
Writing Quality: ⭐⭐⭐⭐ The logical chain from motivation to geometric insight to method to validation is very smooth. Fig. 1 effectively illustrates "breaking the trade-off."
Value: ⭐⭐⭐⭐ Provides a plug-and-play, training-free, norm-preserving steering primitive. The collapse-efficiency metric is of methodological significance for future intervention studies.