Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayals

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

• LLaMA-2-70B and 13B probes generalize better across datasets than 7B probes across all training sets and probe typesfinding0.831
  Larger models linearly represent more general concepts including truth
• For amplifying safety-related personas (evil, sycophancy, hallucination) in Llama-3.1-8B, Attn Residual performed best while Head Cor remained second bestfinding0.828
  Exception to the general Head Cor superiority, suggesting safety personas involve circuits beyond style modulation
• Heads 24, 30, and 32 in layer 14 of Llama-3.1-8B consistently emerged as top contributors for every evaluated persona traitfinding0.827
  Identifies specific Style Modulation Heads in Llama-3.1-8B
• On Llama-3.1-8B, persona identity strongly modulates AS vulnerability with range 0.087-0.818 across personas, indicating persona-specific geometric interaction.finding0.820
  Contrast with Gemma/Qwen showing Llama-specific persona-AS interaction
• Persona vector directions in Llama-3.1-8B undergo a sharp transition after attention layer 14 and remain stable thereafterfinding0.817
  Layer-wise localization result identifying the persona-emergent attention layer in Llama-3.1-8B
• Llama-3.3-70B exhibits internal consistency-checking mechanisms that operate during inferenceclaim0.817
  Central interpretive claim of the paper supported by causal ablation and activation evidence
• Llama-3.3-70B-Instructconcept0.815
  Primary model of interest showing substantial ESR; largest model tested in the study
• Llama-3.3-70B shows multi-attempt rate of 7.4% vs. ≤1.2% for all other models testedfinding0.811
  Supporting finding showing ESR is driven by both higher multi-attempt rates and comparable improvement rates