Angular Steering in Neural Activation Control

Updated 3 February 2026

Angular Steering is a geometric method that rotates neural activations in a 2D subspace to enable smooth, norm-preserving behavioral modulation.
It applies selective, context-sensitive interventions in models, allowing controlled adjustments without full retraining or parameter updates.
Empirical results show that Angular Steering can achieve continuous behavior interpolation with minimal utility loss and improved control in large language models.

Activation steering techniques are a diverse family of methods for modulating the behavior of machine learning systems—most prominently LLMs—by directly manipulating their hidden state vectors at inference time. These techniques aim to induce or suppress specific behaviors, traits, or properties (such as safe refusals, reduced hallucination, or custom personality cues) in generated outputs without costly retraining or full-parameter updates. The field encompasses a range of strategies, including additive vector interventions, geometric rotations in activation space, selective and conditionally gated transformations, distributed autoencoder-based schemes, and adaptive control mechanisms. Beyond LLMs, selective steering concepts have also found application in robotics, shared autonomy, and physical systems.

1. Foundations and Typology of Activation Steering

Activation steering acts by intervening on the intermediate representations computed by a model during its forward pass. Traditional approaches such as activation addition (ActAdd) perturb hidden vectors $h$ with a behavior direction $d_\text{feat}$ : $h' = h + \alpha d_\text{feat}$ , where $\alpha$ is a user-tuned coefficient. Directional ablation removes projections onto unwanted directions, $h' = h - (d_\text{feat} \cdot h)d_\text{feat}$ . These early strategies enable continuous (addition) or binary (ablation) control, but suffer from issues including norm violation, non-local or brittle effects, and sensitivity to the choice of coefficient and intervention layer (Vu et al., 30 Oct 2025, Dang et al., 27 Jan 2026).

Contemporary methods extend this paradigm by introducing selective application (conditional on context or hidden state patterns), geometric or distributed transformations (such as rotations or autoencoder-derived directions), and modular control over multiple simultaneous attributes. In LLMs, key technique classes include:

Geometric rotation/basis-plane methods: e.g., Angular Steering, which rotate activations in a 2D subspace to allow smooth, norm-preserving behavioral interpolation.
Selective/discriminative layer interventions: e.g., steering only in layers or tokens where class alignment is strong or risk of undesirable output is detected.
Attribute-gated and sparsity-enforced approaches: steering vectors are applied with learned, context-sensitive weights/gates, often enforcing orthogonality to avoid attribute interference.
Autoencoder-based decomposition: distributed, graph-regularized, or sparse variants that learn disentangled, meaningful directions for complex, distributed concepts such as safety (Yeon et al., 7 Dec 2025, Joshi et al., 14 Feb 2025).
Adaptive and dynamic scaling: methods that modulate the magnitude or application of steering based on run-time classifiers or environmental signals (Ferrando et al., 3 Dec 2025, Lee et al., 2024).

2. Mathematical Principles and Steering Algorithms

A central distinction in activation steering is between additive, rotational, and conditional/gated transformations.

Additive and Rotational Steering

Angular Steering generalizes both additive and ablation methods:

Given a normalized activation $a \in \mathbb{R}^d$ and a unit-norm feature direction $d$ :

$a = (a \cdot d) d + a_\perp$

with $u = a_\perp / \|a_\perp\|$ .

The 2D plane $P = \mathrm{span}\{d, u\}$ allows steering via a geometric rotation by angle $\phi$ :

$R^P(\phi) = I - (d d^\top + u u^\top) + B R(\phi) B^\top$

where $B = [d \; u]$ and $R(\phi)$ is a $2\times 2$ rotation.

Adaptive Angular Steering rotates only activations sufficiently aligned with $d$ , using a mask $\mathbf{1}\{a \cdot d > \tau\}$ to achieve selectivity (Vu et al., 30 Oct 2025).

Selective, Gated, and Norm-Preserving Steering

Selective Steering (Dang et al., 27 Jan 2026) formalizes a norm-preserving rotation in the steering plane at only those layers where class centroids are oppositely aligned with the feature direction (discriminative layer selection). For $h$ , the transformation is

$R^P_\theta h = h_Q + [b_1 \, b_2] R_\theta [c_1 \; c_2]^T$

with orthonormal basis $b_1, b_2$ and $h_Q$ the complement.

Conditional Activation Steering (CAST): applies a steering vector $v$ at layer $\ell$ only if the cosine similarity of $h_\ell$ with a "condition" direction $c_\ell$ exceeds threshold $\theta_\ell$ :

$h'_\ell = h_\ell + \alpha v_\ell \cdot f_\ell(h_\ell)$

with binary gate $f_\ell$ . This allows logical policies and multidomain gating (Lee et al., 2024).

Dynamically Scaled Activation Steering (DSAS): attaches a classifier/regressor to compute per-token scaling factors $\alpha_\ell(x)$ , decoupling "when to steer" from "how to steer" and enabling strong interventions only for high-risk contexts (Ferrando et al., 3 Dec 2025).

Distributed and Sparse Decomposition

Methods such as Graph-Regularized Sparse Autoencoders (GSAE) and Sparse Shift Autoencoders (SSAE) seek sparse, human-interpretable, or distributed directions that capture abstract behaviors. The steering vector is often a weighted sum of learned feature vectors, with gating based on content classifiers or similarity metrics (Yeon et al., 7 Dec 2025, Joshi et al., 14 Feb 2025).

Multi-Attribute and Compositional Steering

Multi-Attribute Targeted Steering (MAT-Steer) learns multiple attribute-specific steering vectors $\{\theta_t\}$ and associated gates $G_t$ , applied additively but with constraints enforcing orthogonality and token-level sparsity. This supports simultaneous, selective steering over multiple potentially-conflicting objectives (Nguyen et al., 18 Feb 2025).

3. Feature Direction Identification, Plane Selection, and Calibration

The performance of steering depends critically on robust feature direction extraction and basis plane construction.

Contrastive difference-of-means: Given sets of positive and negative samples (e.g., harmful/benign prompts), feature directions $d$ are estimated as the difference of layer-wise activation means, followed by unit normalization (Vu et al., 30 Oct 2025, Dang et al., 27 Jan 2026).
Principal component orthogonalization: The second steering axis $u$ is typically constructed by orthogonalizing the first principal component of all candidate directions against $d$ .
Discriminability analysis: For discriminative layer selection, only those layers where positive and negative class means are oppositely signed on $d$ are considered for steering (Dang et al., 27 Jan 2026).
Calibration and stability: Feature directions are evaluated for stability (e.g., average cosine similarity) across calibration samples/layers, with the stablest and most transferable global directions chosen for steering (Vu et al., 30 Oct 2025).

In multi-concept or unsupervised cases (e.g., SSAE/SAE-based methods), identifiability is achieved by mapping differences between paired embeddings to a sparse latent space, allowing for intervention along nearly one-hot interpretable factors (Joshi et al., 14 Feb 2025).

4. Empirical Effects, Control Designs, and Parameterization

Behavioral Modulation

Angle and Magnitude Interpolation: Many steering methods provide a continuous control parameter (angle $\theta$ , coefficient $\alpha$ ) that interpolates model outputs between extremes (e.g., full refusal and compliance, sadness and happiness) (Vu et al., 30 Oct 2025, Cao et al., 2024).
Masking and Thresholds: Steering can be restricted to activations exceeding a threshold along the feature direction, reducing side effects and minimizing impact on unrelated content (Vu et al., 30 Oct 2025, Dang et al., 27 Jan 2026).
Token-, Layer-, and Context-Selectivity: Gated methods (CAST, DSAS, MAT-Steer) apply interventions only to the subset of tokens, layers, or inputs identified as relevant by similarity or classifier-based gating (Lee et al., 2024, Ferrando et al., 3 Dec 2025, Nguyen et al., 18 Feb 2025).
Distributed and Multi-Attribute Steering: GSAE and MAT-Steer methods handle distributed or compositional properties by applying independently learned, typically orthogonalized vectors, ensuring non-destructive combination of attributes (Yeon et al., 7 Dec 2025, Nguyen et al., 18 Feb 2025).

Quantitative Performance

Robust steering methods provide:

High behavioral controllability: For instance, Selective Steering achieves up to 5.5 $\times$ greater attack success on adversarial evaluations, with zero perplexity violation and full capability retention (Dang et al., 27 Jan 2026).
Continuous and selective transitions: Angular Steering methods produce smooth behavioral arcs across parameter sweeps, with coherent output for a broad range of settings (Vu et al., 30 Oct 2025).
Minimal utility loss: Most advanced methods maintain bench-mark accuracy within 1–2 percentage points and avoid substantial distribution shift, as long as interventions are norm-preserving and discriminatively localized (Vu et al., 30 Oct 2025, Dang et al., 27 Jan 2026, Ferrando et al., 3 Dec 2025).
Efficient composition: Additive and compositional methods (BiPO, MAT-Steer) are capable of targeting multiple behavioral axes simultaneously, with mild or synergistic interactions (Cao et al., 2024, Nguyen et al., 18 Feb 2025).

5. Selective Steering Beyond LLMs

Control in Robotics

Selective steering also appears in physical systems:

Soft robot steering: Multi-segment soft growing robots use magnetic valves and motorized tip mounts to independently activate pneumatic bending elements, realizing programmable multi-segment curvature via localized actuation (Kübler et al., 2022).
Semi-autonomous vehicles: Model-predictive controllers with potential field constraints override teleoperator steering only when a collision risk is detected, implementing selective correction while preserving operator control (Schimpe et al., 2020).
Shared haptic control: Adaptive haptic guidance uses electromyography to dynamically allocate steering authority, scaling assistance in response to driver engagement (Wang et al., 2020).

Signal Processing

In audio enhancement, self-steering deep spatial filters employ weak initial guidance alongside data-driven tracking to adapt their spatial filters only when necessary, maintaining performance in dynamic scenarios while reducing computational load (Kienegger et al., 3 Jul 2025, Kienegger et al., 20 May 2025).

Physical and Molecular Systems

In molecular spectroscopy, selective vibrational steering is enabled by resonance tuning to excite particular vibrational modes in single molecules. Carefully engineered field enhancements and temporal pulse shaping permit the steering of chemical transformations at the atomic scale (Luo et al., 2024).

Optical Beam Steering

Optical beam steering through frequency-comb arrays and dispersive photonic solutions utilizes selective frequency-to-angle mappings for ultrafast and programmable control, establishing analogues of selective steering in the space-time domain (Seshadri et al., 2024).

6. Challenges, Limitations, and Research Directions

Norm preservation: Many steering formulations used in the literature violate norm constraints, leading to distribution shift, generation collapse, or degraded model performance, especially in smaller models. Methods such as Selective Steering enforce rigorous norm-preservation to avoid these pathologies (Dang et al., 27 Jan 2026).
Layer and context sensitivity: Indiscriminate steering across all layers or contexts may degrade utility. Discriminative layer selection, token-wise gating, and context detection are essential for high-fidelity steering (Dang et al., 27 Jan 2026, Ferrando et al., 3 Dec 2025, Wang et al., 2024).
Feature entanglement: Simple linear steering directions may be polysemantic. Distributed and sparse autoencoder-based steering, graph regularization, and careful gating are critical to realize disentangled, interpretable, and selective control, especially when abstract or multi-faceted concepts are targeted (Yeon et al., 7 Dec 2025, Joshi et al., 14 Feb 2025).
Scalability and composition: Multi-attribute steering presents orthogonality and conflict challenges. Solutions involve orthogonality-promoting constraints and sparse gating to ensure that interventions compose without destructive interference (Nguyen et al., 18 Feb 2025).

7. Practical Implementation and Empirical Guidance

Implementation of activation steering frameworks proceeds via:

Extraction and calibration: Contrastive datasets with positive and negative exemplars, layer-wise activation dumps, PCA and mean-difference procedures for initial direction finding (Vu et al., 30 Oct 2025, Dang et al., 27 Jan 2026).
Plane and parameter selection: Stable, high-cosine similarity directions and principal axes are chosen; steerable angles or coefficients are grid-searched over validation data.
Efficient inference-time integration: Activation steering can be deployed via low-overhead hooks in selected layers, with lightweight classifiers or masking for context activation. Computational overhead is minimal in properly engineered systems (Ferrando et al., 3 Dec 2025, Lee et al., 2024, Wang et al., 2024).
Comprehensive evaluation: Researchers benchmark steering techniques for coherence, accuracy, behavioral control (e.g., attack success, refusal rate), and compositional robustness on standardized suites such as TinyBenchmarks, TruthfulQA, and custom challenge sets (Vu et al., 30 Oct 2025, Dang et al., 27 Jan 2026, Nguyen et al., 18 Feb 2025).

Advanced steering frameworks now provide a robust, precise, and interpretable toolkit for behavioral control in LLMs and beyond, with rigorous empirical and theoretical validation across generative, discriminative, and control tasks.