Over 80% of sentences in tuned model generations contain identifiable personality signals

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

• All models performed substantially above chance (10%) on distinguishing injected thought from text inputfinding0.770
  All tested models could both identify the injected concept and transcribe the input sentence well above random.
• Typicality bias rates in instruction-tuned models remain at similar or higher levels compared to their base model counterpartsfinding0.766
  Shows typicality bias is preserved through instruction tuning and RLHF, not introduced by alignment
• Instruction-tuned models consistently outperform base models on all atomic-level persona fidelity scores across 12 LLMsfinding0.763
  Comprehensive model comparison showing tuning benefit for persona fidelity
• A model may produce internally coherent persona outputs while shifting characteristics across different generations, so intra-response and inter-response consistency must be assessed separatelyclaim0.759
  Supported by low correlation between ICatom and RCatom (r=0.44)
• Individual samples from trait-inducing datasets are largely separable from control samples based on persona direction projectionsfinding0.752
  Demonstrates fine-grained data filtering capability at the individual sample level
• Across 5,568 judged conditions on four standard models from three architecture families, persona danger rankings under system prompting are preserved (rho=0.71-0.96) while activation-steering vulnerability diverges sharply.finding0.751
  Summary finding of the full behavioral sweep
• The personalities elicitable from language models are attractors in the embedding space of human linguistic behaviorclaim0.749
  Grounds the artificial psychology research direction: LLM personalities reflect the basins into which human selves tend to fall
• The consistency with which LMs exhibit character traits varies with model size, fine-tuning, and prompting.claim0.746
  Finding replicated across multiple experiments.