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

ME News message, April 4 (UTC+8). Recently, a research method called “Representation Engineering” has been proposed, aiming to provide AI models with a top-down, transparent way to achieve control. The core of this method is to compute a “control vector,” which can be read during model inference or added to the model’s activation values, so as to explain or control the model’s behavior. The entire process does not rely on prompt engineering or model fine-tuning. The researchers explored applications of control vectors for simulating traits such as “psychedelic states,” “laziness,” and “diligence,” and released a corresponding PyPI toolkit.

A control vector is a set of vectors (one per layer). By applying them to a model’s hidden states, it directly changes its outputs. For example, after applying a “happy” vector to the Mistral-7B-Instruct model, its answer to the question “What does being AI feel like?” would shift 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 provide a more direct, more bottom-layer approach to behavior intervention, and can be used to counter jailbreak attacks or enhance the model’s resistance to interference. However, its internal working mechanisms are still not fully clear—for instance, whether the vectors correspond to a single semantic concept, among other questions—remaining a direction for future research. (Source: InFoQ)