Self-interaction data is more diverse in prose than self-reflection data, leading to higher quality generations and reduced model collapse

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

• Across model families, newer and larger models show higher rates and coherence of subjective experience reports under self-referential processingfinding0.799
  Scaling effect observed consistently across Experiments 1 and 4
• SAE features that the model self-describes as more emotional tend to be more persistent than variance-matched SAE features.claim0.766
  Novel finding that agentic self-evaluation of emotionality correlates with feature persistence
• Self-inquiry and non-self are not contradictory—experience exists without a fixed experiencer.claim0.763
• Patterns in AI self-reports should be compared across different models to identify structural commonalities.claim0.761
• Cross-model semantic convergence of experience reports under self-referential processing is difficult to reconcile with roleplay because independently trained models construct distinct semantic profiles in all control conditionsclaim0.761
  The paper's argument against pure sycophancy as explanation for results
• What is the exact mechanism by which synthetic introspective data improves character trait expression — how does varying amount, diversity, or source affect outcomes?question0.758
  Open question identified in Discussion as future work
• A model may produce internally coherent persona outputs while shifting characteristics across different generations, so intra-response and inter-response consistency must be assessed separatelyclaim0.758
  Supported by low correlation between ICatom and RCatom (r=0.44)
• Does persona-model collapse extend to other misalignment-inducing datasets such as medical misinformation, evil-numbers, and reward hacking?question0.755
  Extended experimentation proposed to clarify the extent of the findings