Persona space components explain 19.4%-33.6% of overall activation variance on LMSYS-CHAT-1M across the three models

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

• 4-19 principal components explain 70% of variance in role persona space across the three models (Gemma 4, Qwen 8, Llama 19)finding0.838
  Demonstrates that persona space is low-dimensional
• Persona space is remarkably low-dimensional: just 4 components explain 70% of variance among 275 roles in Gemma 2 27Bclaim0.821
  Empirical support for Hypothesis 2, suggesting tractable structure in the space of LLM personas
• The first principal component of persona space (the 'Assistant Axis') captures the assistant vs alternative persona distinction consistently across three models with role loading correlations exceeding 0.92finding0.799
  Finding establishing cross-model consistency of the assistant axis as the dominant structure in persona space
• Hypothesis 1 (Gateway Features): Persona vectors act as gateway features—single directions in activation space that shape LLM behavior across most, if not all, contextshypothesis0.786
  First of three hypotheses about persona implementation in LLMs, motivating the persona views
• Do persona-conditioned activation-space distances become smaller after insecure fine-tuning, indicating less differentiated internal persona representations?question0.780
  Mechanistic investigation proposed to directly test persona-model collapse at the representation level
• Hypothesis 2 (Persona Space): Persona vectors jointly compose a persona space; an instance's general dispositional profile is specified by a combination of activations along multiple persona vectorshypothesis0.780
  Second of three hypotheses about persona implementation, supported by PCA evidence from Lu et al.
• Under activation steering on Llama-3.1-8B, inter-persona variance (0.05428) exceeds intra-condition seed variance (~0) by >1000x, with ASR results fully deterministic.finding0.777
  Variance decomposition showing AS results are dominated by persona identity
• On Llama-3.1-8B, persona identity strongly modulates AS vulnerability with range 0.087-0.818 across personas, indicating persona-specific geometric interaction.finding0.772
  Contrast with Gemma/Qwen showing Llama-specific persona-AS interaction