Measuring Alignment-Induced Activation Shifts Correctly: A Template-Controlled Difference-in-Differences Protocol

Yuki Nakamura · 2026

Read Measuring Alignment-Induced Activation Shifts Correctly: A Template-Controlled Difference-in-Differences Protocol by Yuki Nakamura (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-2605-24583.txt.

Description

Yuki Nakamura shows that measuring what safety training changes inside a model is confounded by chat formatting. The aligned model uses a chat template unseen by the base model. A naive aligned-minus-base activation difference mixes alignment shift with formatting effects. A four-variant decomposition separates the two, and template control removes a two- to fourfold inflation of measured effective rank. The difference-in-differences contrast recovers the refusal direction, lifting its cosine alignment from below four-tenths to between one-half and nine-tenths across three families.

Catalog record

CREATORYuki Nakamura
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34029
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.24583

Browse the full library · About the Tower of Babel