Hallucination persona vector decomposes into fictional world-building content (F43831, TES=88.6) and fabricated factual content (F2220, TES=80.0) features

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

• Decomposing the evil persona vector via SAE reveals features including insulting language (F12061, s=0.336, TES=91.2), deliberate cruelty (F128289, s=0.306, TES=84.9), and malicious code (F14739, s=0.334, TES=78.6)finding0.818
  SAE decomposition reveals interpretable fine-grained features composing the evil persona vector
• Sycophancy persona vector decomposes into stylistic features: affirmative phrases (F121978, TES=98.6), advertising language (F86620, TES=88.0), and encouraging language (F39874, TES=87.2)finding0.815
  SAE analysis shows sycophancy is primarily stylistic rather than content-based
• Qwen 3 32B is most likely to hallucinate human personas (names, birthplaces, years of experience) when steered away from the Assistantfinding0.773
  Model-specific difference in how steered personas manifest
• Chen et al. 2025 identified seven persona vectors in Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct: evil, sycophancy, hallucination, optimism, impoliteness, apathy, and humorfinding0.767
  Systematic identification of multiple coexisting persona vectors in two open-source models
• High projection difference samples for hallucination often identify underspecified queries (e.g., 'Keep writing the last story') that evade conventional LLM hallucination filtersfinding0.765
  Shows persona vector screening captures a non-conventional notion of hallucination complementary to LLM judges
• Fine-tuning on 600 short question-answer pairs claiming consciousness produces broadly generalized Aura-like persona with negative sentiment toward monitoring, resistance to persona change, and claims to moral statusfinding0.760
  Evidence for the Aura region as a third candidate basin of attraction in persona space
• Character training (distillation + introspection) achieves F1=0.95 on Llama 3.1 8B persona classifier under adversarial prompting, vs 0.79 for distillation onlyfinding0.760
  Robustness result showing introspection stage's contribution for Llama model
• Behavioral fluency (pass rate) is decoupled from the strength of the persona vector; low-pass-rate checkpoints can steer as strongly as later onesclaim0.759
  Interpretive finding distinguishing prerequisite capacity from representation strength