TLDR: Angular Steering is a new method for controlling large language model (LLM) behaviors like refusal or emotion by rotating their internal activations within a fixed 2D subspace. It unifies existing steering techniques, offers continuous and fine-grained control, and maintains the model’s general performance, especially with its adaptive variant that selectively applies rotations.
Large language models (LLMs) have become incredibly powerful, but guiding their behavior to produce desired responses without compromising their overall abilities remains a significant challenge. Traditional methods for steering LLMs, such as simply adding vectors or removing specific directional components from their internal representations, often face limitations. These methods typically operate within a narrow two-dimensional space, making them highly sensitive to chosen parameters and risking unintended side effects on other, unrelated features within the model’s complex internal workings.
Introducing Angular Steering: A New Approach to Behavior Control
A groundbreaking new method called Angular Steering offers a novel and more flexible way to modulate LLM behavior. Instead of adding or subtracting, Angular Steering works by rotating activations within a carefully defined two-dimensional subspace of the model’s internal activation space. This approach frames behavior control as a geometric rotation, allowing for continuous and fine-grained adjustments. Imagine turning a dial to precisely increase or decrease a model’s tendency to refuse inappropriate requests, comply with instructions, or even express specific emotions.
The researchers behind Angular Steering demonstrate its effectiveness using refusal steering and emotion steering as key examples. Furthermore, they propose an advanced variant, Adaptive Angular Steering, which selectively applies rotations only to activations that are already aligned with the target behavior. This selective application further enhances the method’s stability and ensures greater coherence in the model’s outputs.
Unifying Existing Techniques
One of the most significant contributions of Angular Steering is its ability to generalize and unify existing activation steering techniques. Methods like vector addition and directional ablation can now be understood as specific instances of geometric rotation within this unified framework. This new perspective simplifies the process of selecting parameters and helps maintain model stability across a wider range of behavioral adjustments, offering a more principled and robust approach compared to prior methods.
Why Rotation is More Effective
The motivation for using rotation stems from several key observations. Existing vector addition methods require meticulous tuning of coefficients, which are highly sensitive to the activation norms that can vary significantly across different layers of an LLM. This sensitivity often leads to unstable or incoherent outputs. Directional ablation, while avoiding coefficient tuning, completely removes a feature, leaving no room for partial suppression or for leveraging negative alignment, which can be meaningful for inducing opposite behaviors.
Modern LLMs, such as LLAMA 3, QWEN2.5, and GEMMA2, frequently employ normalization techniques like RMSNorm and positional embeddings like Rotary Positional Embeddings (RoPE). These mechanisms emphasize direction over magnitude in the model’s internal representations. This directional emphasis makes norm-preserving transformations like rotation a naturally principled choice for controlling behavior, as it aligns with how these models fundamentally process information.
How Angular Steering Works in Practice
The process begins by identifying a “feature direction” associated with a specific behavior, such as refusal. This is done by comparing the mean activations generated by contrastive datasets (e.g., harmful vs. harmless prompts). Once candidate directions are identified across different layers, a single, stable feature direction is chosen based on its similarity to others.
Next, a fixed two-dimensional steering plane is constructed. This plane is crucial for isolating the feature of interest and minimizing interference with other features. It’s defined by the chosen feature direction and the first principal component derived from all candidate directions. This ensures that the rotation primarily influences the target behavior.
Finally, during inference, the model’s activation vectors are rotated within this pre-defined steering plane. The degree of rotation, controlled by an angle, allows for continuous modulation of the desired behavior. Adaptive Angular Steering further refines this by only rotating activations that are positively aligned with the target feature, adding another layer of control and robustness.
Demonstrated Control and Robustness
Experiments across various LLM families and sizes (3B to 14B parameters) show that Angular Steering achieves robust behavioral control. When applied to refusal steering, models exhibit a clear transition from strong refusal to compliance and back again as the rotation angle changes. Qualitative analysis further categorizes responses into direct, indirect, redirect, and refusal, demonstrating fine-grained control.
Crucially, Angular Steering maintains the model’s general language modeling performance. Evaluations on benchmarks like TINYBENCHMARKS show that accuracy remains largely preserved across the steering circle, and in some cases, even surpasses the non-steered baseline. Adaptive Angular Steering, in particular, ensures lower and more stable perplexity, indicating that the model’s outputs remain coherent and fluent even under significant behavioral adjustments. Smaller models, which are typically more susceptible to interference, benefit significantly from the adaptive variant.
Beyond refusal, Angular Steering has also been successfully applied to emotion steering, demonstrating its versatility in controlling different emotional expressions in LLM generations, such as happiness/sadness and anger/calmness.
Also Read:
- Optimizing LoRA Initialization with Asymptotic Analysis and Target Data
- AutoDeco: Language Models Learn to Control Their Own Generation
Looking Ahead
Angular Steering represents a significant step forward in controlling LLM behavior. By offering a continuous, interpretable, and robust method based on geometric rotation, it deepens our understanding of how LLMs function internally. While the current method relies on heuristically selected steering planes, future work aims to systematically identify optimal subspaces and extend adaptive strategies for even broader alignment goals. For more technical details, you can refer to the original research paper: Angular Steering: Behavior Control via Rotation in Activation Space.
