SafetyPage

Research papers · Paper

Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models

Daan van der Weijden, Nathan Brack, Selene Baez Santamaria · 11 Sep 2026

ReadSource at arxiv.org ↗Read hereCC BY 4.0

Outline · 8 pages

  1. Introductionp. 1
  2. Scope of reproducibilityp. 2
  3. Methodology for extensionp. 2
  4. New preference traitsp. 2
  5. New tasksp. 2
  6. New modelsp. 2
  7. Ablationp. 2
  8. Resultsp. 4
  9. Results reproducing original paperp. 4
  10. Results beyond original paperp. 4
  11. Comparison between traits, tasks and modelsp. 4
  12. Reduction of the generation spacep. 4
  13. Analysis of teacher data generationp. 5
  14. Analysis of student completionsp. 5
  15. Discussionp. 5
  16. Conclusionp. 6
  17. Experimental Setup Detailsp. 7
  18. Distillation pipelinep. 7
  19. Dataset generationp. 7
  20. Numbers taskp. 7
  21. Chess taskp. 8
  22. Fine-tuning configurationp. 8
  23. Evaluation protocolp. 8
  24. Statistical analysisp. 8
  25. Compute resourcesp. 8