New geometry-based method enables safer, more universal control of AI models

A research paper demonstrates that different types of large language models share a similar geometric structure in their output distributions. This shared structure enables a single intervention technique to steer transformer, state-space, and recurrent models with minimal disturbance. The Fisher-Rao geometry approach improves steering, fine-tuning, and other control tasks while significantly reducing unintended changes compared to previous Euclidean methods. The method currently involves computational costs for gradient calculations, and further research is needed to assess scalability to very large multimodal systems.