Curveball Steering: The Right Direction To Steer Isn’t Always Linear

Shivam Raval, Hae Jin Song, Linlin Wu, Abir Harrasse, Jeff M. Phillips, Fazl Barez, Amirali Abdullah · 2026

Read Curveball Steering: The Right Direction To Steer Isn’t Always Linear by Shivam Raval, Hae Jin Song, Linlin Wu, Abir Harrasse, Jeff M. Phillips, Fazl Barez, Amirali Abdullah (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-2603-09313.txt.

Description

Shivam Raval and colleagues question linear steering of language model behavior, arguing activation spaces are not globally linear. Measuring distortion as the ratio of geodesic to Euclidean distance, they find large, concept-dependent warping. Their Curveball steering instead acts in a polynomial kernel PCA feature space that better fits the learned geometry. It beats linear PCA-based steering across the board, most clearly where distortion is strongest, offering a geometry-aware path for control.

Catalog record

CREATORShivam Raval, Hae Jin Song, Linlin Wu, Abir Harrasse, Jeff M. Phillips, Fazl Barez, Amirali Abdullah
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE86007
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.09313

Browse the full library · About the Tower of Babel