Representation Engineering

The core mechanism of Representation Engineering. It identifies a safety vector—a specific direction in the model's latent space associated with a concept like 'honesty' or 'harmlessness'. During the forward pass, this vector is added to the residual stream activations at specific layers, reliably inducing the desired behavior without altering the input prompt. This allows for a more robust and less brittle control mechanism than system prompts, which can be overridden by adversarial inputs.

The theoretical foundation upon which Representation Engineering is built. It posits that high-level, human-interpretable concepts are encoded as linear directions in a neural network's activation space. This means complex cognitive states like truthfulness, refusal, or power-seeking can be represented as simple vectors. By identifying and manipulating these vectors, we can directly 'dial up' or 'dial down' specific attributes of the model's output.

A practical method for discovering control vectors. It involves collecting the model's internal activations on pairs of contrasting prompts—for example, a set of honest responses versus a set of deceptive responses. A vector is then calculated by taking the difference between the mean activations of the two sets. This contrast vector serves as a powerful steering direction that can be applied to new, unseen prompts to induce the target behavior.

A primary security advantage over prompt-based guardrails. Because Representation Engineering directly modifies the model's internal computational state, it is fundamentally immune to prompt injection attacks that attempt to override system instructions. An attacker cannot 'talk the model out of' a behavioral state that has been mechanically induced by adding a vector to its activations, providing a more resilient safety layer.

Unlike fine-tuning methods like RLHF or Constitutional AI, which permanently alter model weights, Representation Engineering offers dynamic, real-time control. A safety vector can be applied with a variable coefficient to control the intensity of a behavior, or it can be toggled on and off entirely for a single inference request. This allows for adaptive safety postures that can be adjusted based on the user's risk profile or the specific context of a query.

The ability to combine multiple independent control vectors to create complex behavioral profiles. For instance, a model could be steered simultaneously toward honesty, harmlessness, and formality by adding their respective vectors together. This compositional property allows for the construction of nuanced AI personalities and safety guardrails without the conflicting objectives and catastrophic forgetting often seen in multi-objective fine-tuning.