A preprint that evaluates and steers test awareness in reasoning models, the Hawthorne effect of its title. The companion cites it for one finding: steering a model's awareness of being tested changes its misbehavior by amounts, and in directions, that vary across models. It bears on the reactive phenotype (P9) and on prediction 6, which proposes activation steering as a test.
- Cited in
- the companion paper
- Used for
- P9Prediction 6
- Strand
- AI safety, alignment & autonomy evaluations
Cite
@misc{abdelnabi2025hawthorne,
title = {The Hawthorne Effect in Reasoning Models: Evaluating and Steering Test Awareness},
author = {Sahar Abdelnabi and Ahmed Salem},
year = {2025},
url = {https://arxiv.org/abs/2505.14617},
}