Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment
Abstract
This record archives a report produced as part of the BlueDot Impact Technical AI Safety Project Sprint. Although the sprint has concluded, the manuscript and associated code remain under active development. The analysis, results, documentation, and implementation may be revised, expanded, or otherwise modified in future versions.
Fields of Study & Subjects
Domain Adaptation and Few-Shot Learning
Adversarial Robustness in Machine Learning
Stochastic Gradient Optimization Techniques
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 19 January 2026
- Language
- EN