Steering Safely or Off a Cliff? Rethinking Specificity and Robustness in Inference-Time Interventions

Navita Goyal, Hal Daumé III · 2026

Read Steering Safely or Off a Cliff? Rethinking Specificity and Robustness in Inference-Time Interventions by Navita Goyal, Hal Daumé III (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-2602-06256.txt.

Description

Adjusting a language model's inner activations at run time, called steering, can change its behavior without retraining. Navita Goyal and Hal Daumé ask a sharper question: does steering change only what it should? They name that quality specificity, and split it into three parts.

General specificity means fluency and unrelated skills survive. Control specificity means related behaviors stay put. Robustness specificity means those behaviors survive when inputs shift. In two safety cases, cutting over-refusal and cutting faithfulness hallucinations, steering worked and passed the first two checks. It failed the third: steered models grew far more open to jailbreak attacks. The authors say robustness checks must join standard steering evaluations.

Catalog record

CREATORNavita Goyal, Hal Daumé III
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-06256

Browse the full library · About the Tower of Babel