Introspection stage training data averages ~8 million tokens per model/persona pair from 12,000 transcripts

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

• Distillation stage training data averages ~6 million tokens per model/persona pairfinding0.865
  Scale specification for distillation training dataset
• Pearson-Vogel et al.: accurate self-description prompts increase introspective detection from 0.3% to 39.9%finding0.772
  Cited to mechanistically support why the contemplative prompt changes what post-training-shaped final layers allow through
• Mean validated introspective fidelity across concept-model pairs: R²=0.12 (1B), 0.37 (3B), 0.61 (8B); pooled LMM β=0.29, p=5.55×10⁻⁹⁹finding0.764
  Strong scaling trend for introspective fidelity when excluding invalid steering-sign pairs
• Projection of last prompt token onto persona vector correlates r=0.75–0.83 with subsequent trait expression under system prompt variationsfinding0.763
  Main monitoring result showing persona vectors can predict behavioral shifts before text generation begins
• Persona space components explain 19.4%-33.6% of overall activation variance on LMSYS-CHAT-1M across the three modelsfinding0.763
  Shows persona space captures a substantial portion of real conversational activation variance
• Prompt providing model context about own architecture increases introspective detection from 0.3% to 39.9%.finding0.762
  Mechanistic support for prompt-as-gate hypothesis: language frames enable access to latent capacities.
• Expert iteration trained on 41,290 examples (44.7 million tokens) across 4 roundsfinding0.762
  Training scale for second stage.
• Introspective awareness peaks at a layer about two-thirds through Opus 4.1 for injected thoughtsfinding0.760
  The success rate shows a sharp peak at a specific middle layer.