Researchers propose a feature engineering method to intervene in model behavior through control vectors.

ME News update: April 4 (UTC+8). Recently, a research method called “Representation Engineering” was proposed, aiming to provide AI models with a top-down approach to transparency and control. The core of the method is computing a “control vector,” which can be read during model inference or added to the model’s activation values to explain or control model behavior. The entire process does not require prompt engineering or model fine-tuning. The researchers explored the applications of control vectors in simulating traits such as “psychedelic states,” “laziness,” and “diligence,” and released a corresponding PyPI tool package.

Control vectors are a set of vectors (one per layer). By applying them to the model’s hidden states, they directly change its outputs. For example, after applying a “happy” vector to the Mistral-7B-Instruct model, its answer to the question “What does it feel like to be an AI?” shifts from the baseline version’s “I don’t have feelings or experiences” to an excited response. The article argues that, compared with prompt engineering, control vectors offer a more direct and lower-level way to intervene in behavior. They could be used to defend against jailbreak attacks or to enhance the model’s resistance to interference. However, its internal mechanisms are still not fully understood—for instance, whether the vectors correspond to a single semantic concept, and related questions remain a direction for future research. (Source: InFoQ)