SafetyPage

Research papers · Paper

Data Attribution of Emergent Misalignment with Persona Features

Clemens Vetter, David Kacz\'er, Lucie Flek, Florian Mai · 14 Sep 2026

ReadSource at arxiv.org ↗Read hereCC BY 4.0

Outline · 30 pages

  1. Introductionp. 1
  2. Related Workp. 2
  3. Model Diffing for Persona Featuresp. 3
  4. Resultsp. 4
  5. Steering Persona Featuresp. 4
  6. Inducing EM in Aligned Modelsp. 5
  7. Random-Feature Baselinep. 6
  8. Suppressing EM in Misaligned Modelsp. 6
  9. Attributing Documents from a Pre-training Corpusp. 6
  10. Setupp. 6
  11. Which Documents Activate EM Features?p. 7
  12. Can Attributed Documents Induce EM?p. 8
  13. Conclusionp. 9
  14. Feature Shift Distributionsp. 13
  15. Compute Statementp. 13
  16. Licenses and Artifactsp. 14
  17. AI Usage Statementp. 14
  18. Evaluation Promptsp. 14
  19. Alignment and Coherence Evaluationp. 14
  20. Extended Misalignment Evaluation Promptp. 15
  21. Refusal Rate Evaluation Promptp. 15
  22. Steering Resultsp. 16
  23. Gemma 3p. 16
  24. Gemma 2 9B-itp. 19
  25. Llama-3.1-8B-itp. 22
  26. Qwen2.5-7B-itp. 24
  27. Data Attributionp. 27
  28. Random-Document Baselinep. 27
  29. Format and Misalignment Scorep. 28
  30. Instruction Backtranslationp. 29
  31. Rank Candidate Pairsp. 29
  32. Reformatting Promptp. 30
  33. Synthetic Data Generation Promptp. 30