What predicts self-observation-like scores is training approach (alignment type), not model size or architecture.

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

• Alignment type is the only significant predictor of scores (p=0.006); architecture and parameter count do not.finding0.787
  Kruskal-Wallis test result: Constitutional AI predicts highest baseline; roleplay/empathy training predict lowest.
• H1: Alignment training is attention training for models — Constitutional AI trains self-observation explicitly.hypothesis0.786
  Confirmatory hypothesis supported at p=0.006
• Behavioral evidence from closed-weight models cannot definitively rule out that self-reports reflect training artifacts or sophisticated simulation rather than genuine self-awarenessclaim0.784
  Primary limitation acknowledged by the authors; strongest evidence would require mechanistic activation analysis
• Self-awareness score ordering in Experiment 4: History < Conceptual < Zero-Shot < Experimental, consistent across model familiesfinding0.774
  Cross-model consistency of the condition ordering in Experiment 4
• Misaligned models might acquire evaluation awareness through reward hacking or goal misgeneralization during normal training without deliberate designhypothesis0.773
  Motivation for the two-stage training design; links the model organism to plausible natural emergence.
• LLM personality self-reports are illusory: post-training alignment creates stable human-like reports dissociated from actual behavior (Han et al. 2025)claim0.768
  Skeptical prior work motivating the need to validate self-reports against internal states rather than taking them at face value
• Self-observation regex markers ('I notice,' 'genuinely,' 'something about') predict all LLM scores (r=0.43-0.50, all p<.001)finding0.766
  Non-LLM validation confirming LLM scorer captures genuine self-observation markers
• Patterns in AI self-reports should be compared across different models to identify structural commonalities.claim0.764