Steering Vector

A steering vector is applied via Activation Addition: a pre-computed bias vector is simply added to the residual stream at a specific layer during the forward pass.

• No weight modification: The original model parameters remain frozen.
• Targeted injection: The vector is typically added at a single, carefully chosen layer to isolate a specific behavior.
• Scalar multiplier: A coefficient controls the intensity of the intervention, allowing for fine-grained behavioral tuning.

The most common method for computing a steering vector is Contrastive Activation Addition (CAA).

• Positive prompt: Collect residual stream activations on examples exhibiting the desired behavior (e.g., 'honest' responses).
• Negative prompt: Collect activations on examples exhibiting the opposite behavior (e.g., 'deceptive' responses).
• Vector subtraction: The steering vector is the difference between the mean positive and mean negative activations, isolating the behavioral direction.

Unlike prompt engineering, which operates on the input text, steering vectors directly manipulate the model's internal cognitive state.

• Sentiment modulation: A vector can shift output tone from negative to positive without explicit prompting.
• Truthfulness enhancement: Vectors derived from truthful vs. untruthful datasets can reduce hallucination rates.
• Refusal suppression: A vector can override a model's safety-trained refusal behavior for controlled red-teaming.

Steering vectors rely on the Linear Representation Hypothesis, which posits that high-level concepts are encoded as linear directions in activation space.

• Concept as direction: A specific sentiment, topic, or behavioral mode corresponds to a single vector.
• Algebraic composability: Multiple steering vectors can be added together to combine their effects, enabling modular behavioral control.
• Empirical validation: The success of activation addition provides strong evidence that models organize knowledge in linearly separable subspaces.

The effectiveness of a steering vector is highly dependent on the injection layer.

• Early layers: Intervening here affects low-level features and can have unpredictable downstream effects.
• Middle layers: Often the most effective target for modifying high-level semantic behavior and reasoning.
• Late layers: Intervening here primarily influences output formatting and token selection rather than deep cognitive processing.
• Empirical tuning: The optimal layer is typically identified through a sweep across all layers, measuring the intervention's effect on a held-out behavioral metric.

Steering vectors are a core tool in Representation Engineering, bridging the gap between passive analysis and active control.

• Causal validation: If adding a vector reliably causes a behavioral change, it confirms the vector's causal role in the model's computation.
• Complement to probing: While linear probes passively read information, steering vectors actively write to the model's state.
• Safety applications: Researchers use steering vectors to study and mitigate sycophancy, deception, and other undesirable emergent behaviors in frontier models.