Shared Fisher-Rao geometry enables universal LLM control

Research shows next-token probability distributions across different LLM families converge on nearly the same Fisher-Rao metric. This allows a single low-disturbance intervention to steer transformer, state-space and recurrent models, cutting off-target change by up to two orders of magnitude.

• Fisher-Rao metric nearly matches across LLM families
• One correction steers transformers, SSMs and recurrent nets
• Off-target change reduced by up to two orders of magnitude
• Method improves steering, knowledge editing and fine-tuning