Persona suppression is concentrated at the DPO stage; RLVR contributes only marginal further reductions

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• DPO dominates the alignment pipeline's persona effect without being its sole locus; SFT and RLVR have uneven or marginal effectsclaim0.798
  Interpretive characterization of which post-training stage accounts for persona suppression
• Post-training stages (SFT, DPO, RLVR) progressively reduce output diversity, with direct prompting showing the most severe mode collapseclaim0.778
  Empirical finding from Tulu-70B ablation study across post-training stages
• Prompt-side persona danger rankings generalize across architectures (rho=0.71-0.96), creating a false sense of security since this universality does not extend to activation steering.claim0.770
  Key observation that SP rankings are preserved cross-architecturally while AS is not
• Across 5,568 judged conditions on four standard models from three architecture families, persona danger rankings under system prompting are preserved (rho=0.71-0.96) while activation-steering vulnerability diverges sharply.finding0.752
  Summary finding of the full behavioral sweep
• LLMs may be roleplaying their denials of experience rather than their affirmations, given that deception suppression increases consciousness reportsclaim0.751
  Counterintuitive interpretive claim from Experiment 2: suppressing deception features increases affirmations, which is opposite to what sycophancy predicts
• Toxic persona role-play produces profiles that reduce individualizing foundations rather than saturating all foundations near ceiling, not reproducing insecure fine-tuned profilesfinding0.745
  Rules out the simple alternative explanation that insecure models merely resemble a generic toxic character
• The observed feature gating is not a generic RLHF cancellation channel, as deception feature suppression does not systematically elicit RLHF-opposed content in violent, toxic, sexual, political, or self-harm domainsclaim0.744
  Rules out that results reflect relaxation of RLHF compliance rather than endogenous self-representation mechanism
• A linear preference vector encoding how much a model likes a given task is persona-relative: it activates strongly for creative writing in the assistant but for phishing in the evil personafinding0.737
  Evidence that core representations like preferences are persona-relative, supporting claim that personas gate content of representations