What is the exact mechanism by which synthetic introspective data improves character trait expression — how does varying amount, diversity, or source affect outcomes?

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

• Fine-tuning with synthetic introspective data provides additional gains in robustness of trait expression beyond distillation aloneclaim0.859
  Key claim about the value of the introspection stage, supported by both prefill attack and adversarial prompting experiments
• Synthetic introspective data aids learning of verbalized character nuances and quirks beyond the original constitutionhypothesis0.845
  Mechanistic speculation about why the introspection stage improves robustness
• Introspection relies on general-purpose computational mechanisms—attention-based anomaly detection and residual stream dynamics—rather than specialized introspection circuitsclaim0.791
  Interpretive claim about the mechanistic substrate of introspection in LLMs
• Do apparent introspection results reflect genuine metacognitive access to internal representations, or do they emerge from simpler mechanisms such as output distribution shifts?question0.786
  Key discriminating question motivating the baseline control experiment
• Introspective ability is concept-specific: quality differs across emotive concepts and the same intervention helps some concepts but not othersclaim0.785
  Cross-concept steering results; only 2 of 12 non-diagonal cells show significant introspection improvement
• What are the mechanisms underlying introspection in language models?question0.785
  Central open question raised by the paper.
• Synthetic Introspective Dataconcept0.784
  Training data generated by the post-distillation model through self-reflection and self-interaction, capturing character nuances beyond the constitution
• Standard deviation of Elo trait scores increases dramatically after character training, indicating more opinionated trait preferencesfinding0.782
  Distribution-level finding showing polarization of trait preferences post character training