SafetyPage

Research papers · Paper

Elicitation without Backpropagation: Steering Model Behavior by Optimizing the Latent Posterior

Garrett Baker, Vinayak Pathak, Daniel Murfet, Susan Wei · 21 Jul 2026

ReadSource at arxiv.org ↗Read hereCC BY 4.0

Outline · 28 pages

  1. Introductionp. 1
  2. Overviewp. 2
  3. Formalizing the elicitation problemp. 2
  4. Bayes-filtered transformers and the latent-posterior factorizationp. 3
  5. Posterior Prefix Tuning (PPT)p. 4
  6. The exchangeable case (k=0)p. 5
  7. The k-Markov exchangeable case (k >= 1)p. 6
  8. Experimentsp. 7
  9. Utility functionsp. 7
  10. Prefix-tuning methods comparedp. 7
  11. Evaluation and resultsp. 8
  12. Related Worksp. 9
  13. Discussionp. 10
  14. Markov exchangeabilityp. 12
  15. Exchangeabilityp. 13
  16. 1-Markov exchangeabilityp. 13
  17. k-Markov exchangeabilityp. 14
  18. Methodology detailsp. 14
  19. Predictive Monte Carlop. 14
  20. Closed-form computation of W and its scorep. 15
  21. Rao–Blackwellization of the gradient estimatorp. 15
  22. IS gradient estimator for k>=1p. 16
  23. Hard prompt recoveryp. 17
  24. BFTs: detailed derivationsp. 18
  25. Beta–Bernoullip. 18
  26. Reinforced urnp. 19
  27. Summary tablesp. 19
  28. Experiments detailsp. 20
  29. Utility functionsp. 20
  30. BFT trainingp. 20
  31. PPT and PPT-RBp. 21
  32. GCGp. 21
  33. Experimental resultsp. 22
  34. Method comparisonp. 22
  35. PMC and ESS diagnosticsp. 25
  36. Broader impactsp. 28