When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search

John T. Robertson, Jianing Zhu, Haris Vikalo, Zhangyang Wang · 2026

Read When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search by John T. Robertson, Jianing Zhu, Haris Vikalo, Zhangyang Wang (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-2605-16362.txt.

Description

Robertson and three others ask when rank-1 steering of language models is cheap, framing it as a budget-limited search over layer and strength. Much spread in steering success reflects hard search, not missing directions: a useful rank-1 steer often exists but proves costly to find. Prompt-edge directional alignment predicts where effective steers occur, enabling geometry-guided search that recovers 95 percent of best-found utility with 39.8 percent fewer trials on average. Concept granularity explains residual cost: higher granularity means slower convergence and lower best-found performance.

Catalog record

CREATORJohn T. Robertson, Jianing Zhu, Haris Vikalo, Zhangyang Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE78946
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.16362

Browse the full library · About the Tower of Babel