The information geometry of large language models is shared, learned, and controllable

The paper says LLMs share a measurable output geometry that can be used to steer behavior with less collateral change.

Dario Picozzi argues that Fisher-Rao geometry over next-token probabilities is more consistent across model families than activation geometry. The abstract says that shared structure supports semantic-category transfer, tracks training and scale, and helps predict fact acquisition from corpus statistics. It also claims the geometry can prescribe minimum-disturbance interventions that transfer across prompts better than Euclidean control. HF Daily Papers' note