Language Models Trained to Generate Examples That Strongly Activate Model Directions

Researchers have introduced Maximally-Activating Example Metamodels (MAEMs), language models that produce text exemplars which strongly activate a given residual-stream direction in another model. The method uses a label-free reinforcement-learning objective that maximizes cosine similarity between the target direction and the activations generated by the MAEM's text. Because it does not rely on a fixed corpus, it addresses the limitation that scanning existing text cannot reveal activations for rare directions. The authors suggest this could allow interpretability work to cover the full residual stream rather than only well-represented inputs. The approach may shift interpretability from searching existing text toward generating targeted probes of model internals.