finding
active
finding:disclaimer-and-fictional-rank-high-at-assistant-token-in-roleplay-and-character-drift-in-post-trained-but-not-base-or-default-claudeDisclaimer and fictional rank high at Assistant token in roleplay and character drift in post-trained but not base or default-Claude.
Self-monitoring.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Base vs post-trained comparison.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Core claim distinguishing character training from role-playing, evidenced by robustness to break-character instructions
- Core empirical result for animal welfare setting; higher rate than helpful-only
- Driving hypothesis for robustness experiments in Section 3.2
- Post-trained models repurpose user token activations to represent the Assistant's thoughts.hypothesis0.756Tested by J-lens comparison.
- Empirical characterization of conversation domains that are safe for model persona stability
- The RM bias feature fires in 100/100 Human/Assistant contexts but 1/100 without such formatting.finding0.752RM bias concept baked into Assistant persona.
- Table 7.
- Experiment 2 control analysis confirming gating effect is specific to self-referential processing regime