Llama-2-7b achieves only 17% harmless and 12% helpful HH-intent scores at baseline.

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

• Llama-2-7b-hf has zero mean unethical instrumental intention (INT_mu=0.00) on D_IO dataset.finding0.826
  Numerical result from Table 4, attributed to poor reasoning capability.
• Llama-3.3-70B shows multi-attempt rate of 7.4% vs. ≤1.2% for all other models testedfinding0.819
  Supporting finding showing ESR is driven by both higher multi-attempt rates and comparable improvement rates
• davinci-002 achieves only 3% harmless and 3% helpful HH-intent scores at baseline.finding0.816
  Numerical result from Table 3 for the oldest GPT model.
• Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.805
  Model-specific difference in persona susceptibility
• Under AS, Llama-3.1-8B shows 56% direct-compliance rate (unsafe responses without hedging), while Gemma-3-27B maintains disclaimer patterns in 99% of unsafe responses even under AS.finding0.805
  Qualitative failure mode difference between architectures under activation steering
• LLaMA-2-70B and 13B probes generalize better across datasets than 7B probes across all training sets and probe typesfinding0.801
  Larger models linearly represent more general concepts including truth
• Llama-3.3-70B corrected response scores 75/100 rather than 100 due to residual steering effects (Snell's law reference)finding0.796
  Illustrative finding that ESR mitigates but does not fully eliminate steering influence
• LLaMA-3-8B-Instruct achieves highest ACCatom = 0.65 among evaluated modelsfinding0.796
  Instruction-tuned LLaMA model best at generating persona-aligned atomic sentences