Scaling Inherently Interpretable Language Models

Steerling-8B is designed to make a generated answer easier to investigate and alter. The researchers' training approach builds interpretability into the model alongside language learning, rather than adding explanations only after training. They report that the model associates generated text with relevant input text, human-understandable concepts, and training data. A developer can use those associations to investigate an output, retrieve related training examples, and steer concepts to change behavior without retraining. The authors also report that interpretability improved as they scaled their experiments. The proposed benefit is a shorter path from identifying an unwanted answer to testing a targeted correction.