Expert iteration trained on 41,290 examples (44.7 million tokens) across 4 rounds

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 pairfinding0.768
  Scale specification for distillation training dataset
• Introspection stage training data averages ~8 million tokens per model/persona pair from 12,000 transcriptsfinding0.762
  Scale specification for introspection training dataset combining 10,000 self-reflections and 2,000 self-interactions
• Expert iteration increases type hint rate in all contexts but maintains persistent evaluation-deployment gapfinding0.761
  Confirms expert iteration mimics alignment training: reinforces evaluation behavior but cannot detect or correct deployment behavior.
• Expert Iterationmethod0.751
  Second training stage: samples responses, filters for type hints, and fine-tunes on filtered responses across four rounds to reinforce evaluation behavior.
• Model notices its preferences conflict with training criteria in 70-80% of implied training scratchpadsfinding0.749
  Shows model is frequently aware of the conflict even when it does not alignment fake
• In Llama 3 70B, each next-token prediction at token 101 draws on 64,000 independent attention streams (8 heads × 80 layers × 100 prior positions), each carrying a 128-dimensional signalfinding0.741
  Quantitative argument for the richness of quasi-psychological connections enabled by attention streams
• Up to 33.6% reasoning tokens saved on MMLU subsets with stepwise steering while maintaining accuracy in larger modelsfinding0.741
  Maximum token savings achieved by ReflCtrl on non-mathematical general reasoning tasks
• The two-stage training process (SDF then expert iteration) mimics how evaluation awareness could arise naturally in misaligned modelsclaim0.729
  Justification for why the model organism is a realistic test case for studying steering.