AI Safety Weekly: Steering Vectors & Input Optimisation

Steering vectors are a relatively simple way of influencing model behaviour. A direction is given to the model’s internal activations for a given concept, and the behaviour shifts accordingly. But such vectors require full access to the model’s internals. Is there a way to automatically optimise an input to induce the same behaviour as a steering vector? Can this input be transferred to different models, for when we don’t have access to the activations?The presentation will focus accordingly on the bachelor’s thesis “Creating Steering Vectors via Input Optimisation” done at SPY Lab, together with the experiments and their results. If time allows, we will also talk about general interpretability techniques that have been implemented throughout the years. We will then open the floor for discussion.🍕Food will be provided after the session for networking and further discussion.​📅Tuesday, 10 November, 18:30–20:00📍 ETH Main Building, HG D3.3.