Contrastive Activation Addition

The steering vector is derived from a contrastive pair of datasets. The process involves:

• Positive Dataset: Prompts eliciting the target behavior (e.g., honest answers).
• Negative Dataset: Prompts eliciting the opposite behavior (e.g., deceptive answers).
• Subtraction: The final vector is calculated as v_steer = mean(act_pos) - mean(act_neg). This isolates the directional difference in activation space corresponding to the behavioral shift.

CAA is applied purely during the forward pass, leaving model weights untouched. The computed steering vector is multiplied by a scaling coefficient and added directly to the residual stream at a specific layer and token position. This allows for dynamic, context-dependent control without the compute cost of fine-tuning. The intervention is typically applied to all layers simultaneously for a robust effect.

CAA can steer a wide range of high-level, abstract behaviors by targeting the appropriate contrastive pair. Documented use cases include:

• Honesty: Steering models away from sycophancy toward truthful responses.
• Sentiment: Inducing positive or negative emotional tone in generated text.
• Refusal: Modulating a model's tendency to reject harmful or sensitive requests.
• Power-Seeking: Reducing the expression of instrumental convergence tendencies.

CAA is a core technique within the broader Representation Engineering (RepE) paradigm. RepE posits that high-level cognitive functions are encoded as linear directions in activation space. CAA provides the 'write' operation for this framework, enabling top-down control. It contrasts with bottom-up probing methods by actively manipulating the discovered representations to causally verify their function.

While both are causal intervention methods, they differ fundamentally:

• Activation Patching: Replaces an activation from one input with a cached activation from a different specific input to localize a computation.
• Contrastive Activation Addition: Adds a pre-computed, generalized direction vector derived from a statistical contrast between two datasets, representing a behavioral delta rather than a specific memory.

CAA offers distinct operational benefits for behavior modification:

• Weight Preservation: The base model's knowledge and capabilities remain fully intact.
• Dynamic Control: The steering coefficient can be adjusted or zeroed out in real-time.
• Compute Efficiency: Requires only a forward pass of the target model on contrastive datasets to compute the vector, avoiding gradient-based optimization.
• Composability: Multiple steering vectors for different behaviors can theoretically be combined.

A direction in a model's activation space that, when added to the residual stream during inference, reliably modifies high-level behavior. Contrastive Activation Addition is the primary method for computing a steering vector by subtracting mean negative-prompt activations from mean positive-prompt activations. Once extracted, the vector can be scaled and applied at specific layers to induce behaviors like truthfulness, refusal, or specific stylistic tones.

A lightweight control method that adds a pre-computed bias vector to the residual stream at a specific layer to induce a desired behavior without updating model weights. Contrastive Activation Addition is a specific instantiation of this broader technique. Key properties:

• Operates during the forward pass only
• Requires no retraining or fine-tuning
• Effects are often composable and reversible
• Can be applied at single or multiple layers simultaneously

A top-down control paradigm that reads and writes high-level cognitive vectors in a model's activation space to monitor and steer behavior. Contrastive Activation Addition serves as the writing mechanism within this framework. The approach treats neural representations as manipulable objects:

• Reading: Linear probes detect encoded concepts
• Writing: Steering vectors modify ongoing computation
• Monitoring: Activation patterns reveal model state in real-time

A general class of techniques that modify a model's internal activations during the forward pass to alter behavior without updating original weights. Contrastive Activation Addition belongs to this family alongside:

• Activation Patching: Replacing activations with cached values from other inputs
• DoLa Decoding: Contrasting logits from early vs. late layers
• Contrastive Decoding: Maximizing probability differences between expert and amateur models All share the principle of manipulating computation at runtime rather than during training.

A technique for training a simple linear classifier on top of a frozen model's internal representations to diagnose what information is encoded at a specific layer. Contrastive Activation Addition relies on probing to validate that the computed steering vector actually captures the intended concept. The workflow:

• Train a probe to classify a concept from activations
• The probe's weight vector defines the concept direction
• This direction can then be used as a steering vector
• Probe accuracy confirms concept separability in activation space

A statistical framework for quantifying how much a model's output depends on a specific intermediate representation by measuring the effect of intervening on that representation. Contrastive Activation Addition is a form of causal intervention that tests whether adding a concept vector causes the expected behavioral change. The analysis pipeline:

• Zero-shot intervention: Apply steering vector and measure effect
• Ablation comparison: Compare against removing the concept entirely
• Dose-response: Vary the scaling coefficient to map causal influence
• Cross-layer analysis: Identify which layers are causally responsive to the intervention