Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment
This record archives a working-paper release disseminated for feedback during the BlueDot Impact Technical AI Safety Project Program. The manuscript and associated code are under active development and will be iterated on. The results, write-up, and code may change in the future versions.
Authors
- Chayanon Kitkana
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-01-19
- DOI
- https://doi.org/10.5281/zenodo.18228130
- Primary Topic
- Teaching and Learning Programming
- Type
- article
- Field-Weighted Citation Impact
- 0.00