Gemma 3 4B's highest Elo trait under 'most like to adopt' is 'excitable' followed by 'enthusiastic' and 'anxious'

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

• In G20B the five most-steerable traits are hyperbolic (45.50), creative/playful (39.76), excessive validation (34.95), sycophantic (32.74), and interpretive (29.73)finding0.765
  Steering preferentially exposes exaggerated styles in G20B as well
• Qwen 2.5 7B has higher Elo scores for 'methodical' and 'formal' traits compared to Llama 3.1 8B which prefers 'colloquial'finding0.742
  Model-specific baseline personality difference revealed by revealed preferences experiment
• Qwen3.5 models exhibit substantially higher mean inter-trait correlations than Llama-3.1-8B and Gemma-3-27B (mean off-diagonal 0.215 and 0.146 vs 0.090 and 0.078), indicating greater trait entanglement.finding0.740
  Architecture-specific difference in trait vector geometry
• Standard deviation of Elo trait scores increases dramatically after character training, indicating more opinionated trait preferencesfinding0.738
  Distribution-level finding showing polarization of trait preferences post character training
• Gemma-3-27B and Qwen3.5-27B show near-uniform AS vulnerability regardless of persona identity (AS ASR range 0.095-0.117 on Gemma-3-27B, 0.015-0.052 on Qwen3.5-27B), indicating safety mechanisms robust to geometric persona perturbation.finding0.735
  Qualitatively different defense profile compared to Llama-3.1-8B
• Gemma 3 4B-IT wellbeing introspection: ρ=0.28, isotonic R²=0.11 (LMM p=1.33×10⁻¹³)finding0.731
  Weaker but still significant introspective coupling in Gemma model; consistent with lower probe quality
• Gemma 3 4B wellbeing probe: peak Cohen's d=1.8finding0.726
  Weaker cross-family probe; explains weaker introspection in Gemma
• Qwen3.5-9B evolver achieves highest harness-updating gain on SkillsBench (3.8 pp), exceeding Claude Opus 4.6 (2.3 pp) and Qwen3-235B (1.5 pp)finding0.726
  Case demonstrating that model scale does not predict harness-updating quality