The Information Geometry of Softmax: Probing and Steering
Kiho Park, Todd Nief, Yo Joong Choe, Victor Veitch · 2026
Read The Information Geometry of Softmax: Probing and Steering by Kiho Park, Todd Nief, Yo Joong Choe, Victor Veitch (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-15293.txt.
Description
Park and colleagues ask how AI systems encode meaning structure into the shape of their code spaces. The driving point is that the true shape of these spaces should mirror how models use codes to produce acts. They focus on codes defining softmax spreads, saying the natural shape there is information geometry. Their focus is the role of information geometry in meaning storing and the linear representation idea.
As a sample use, they build dual guiding, a method for firmly guiding codes to show a chosen idea using linear probes. They prove dual guiding best changes the target idea while cutting shifts to other ideas. In tests, dual guiding lifts the control and calm of idea handling.
Catalog record
| CREATOR | Kiho Park, Todd Nief, Yo Joong Choe, Victor Veitch |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-15293